diff --git a/CHANGELOG.md b/CHANGELOG.md index b807939f0..67650c887 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -4,6 +4,12 @@ ## [Unreleased] +### Memadapter 论文精读入库：记忆诱发谄媚的检索后治理（cognitive-context 015，2026-10-28） + +- 新增 `docs/research/cognitive-context/015-memadapter-counterfactual-adaptation.md`（arXiv:2610.05162v1，CC BY 4.0）：受控实验证「准确记忆也诱发谄媚」（96.0%→77.3%）→ 反事实归纳任务无关边界（式1–3）→ 任务条件化使用指令（式4）→ 证据锚定生成＋支撑轨迹核验（式5）；三基准五记忆系统三骨干主结果、消融与运行时，五规律三争议与批判边界五条。 +- 配套纯标准库原型 `assets/memadapter_lab.py`（九场景确定性仿真：变体阶梯 3/9→7/9→8/9→9/9，E1–E4 破坏实验实测退化）；archify 三阶段机制图四件套入库（`memadapter--three-stage-pipeline`）；论文原图 4 张（图1/2a/2b/3）入 source-figures 并登记台账。 +- 登记面同步：research readme、knowledge-map、mermaid README。 + ### Daily Stock Analysis（DSA）精读与通俗拆解入库（agent-harness 178，2026-10-27） - 新增 [178-daily-stock-analysis](docs/research/agent-harness/178-daily-stock-analysis.md)：以 67k star 开源仓 ZhuLinsen/daily_stock_analysis 钉点 `4e622643`（2026-10-25，产品站 v3.33.0-10-g4e622643 同源构建）+ 站点 12 任务指南原生页 + 系统论文 arXiv:2608.26990v1 三源首读：五段证据到报告流水线与双执行路径、八态证据状态与六块质量评分（缺数据≠中性、核心块降级禁高置信）、多源回退三层防护（回退链/熔断/冷却）、模型路由三层配置×两层路由与能力准入（可替换≠等价；13 家 prompt-cache 登记未开门的预留态）、15 策略技能与信号资格分选（无效进诊断、互斥穷尽）、分歧显式化+RiskGuard 只降不升、决策信号 1d–10d 后验与 Beta(15,15) 收缩权重自校准；五条底层规律、三个争议与「论文 1457/596 契约测试只证软件一致性」的批判边界五条。 @@ -12,7 +18,6 @@ - 登记面：research readme / knowledge-map / mermaid README / source-figures README / issue I-239。 - 验收：56 行源稿对账（8 行返工复判收口，COUNTS 全零）+ 增量 Checker 核外行补阙 26 行 + 盲评 3 章（成稿全胜）+ 外行四测四项出闸（LLM 代理作答；两处事实错误经测出并修正：信号别名表无中文、hold 官方中文为「持有」）。 -### Harness 到 Harness² 双源精读与通俗拆解入库（agent-harness 179，2026-10-28） ### Harness 到 Harness² 双源精读与通俗拆解入库（agent-harness 179，2026-10-28） - [179 精读](docs/research/agent-harness/179-harness-to-harness2.md)：Harness-Zero（arXiv:2609.24974v1，训练期审阅门蒸馏进权重、部署卸载）× Harness²（arXiv:2610.14224v1，部署期冻结模型的对比式 PPC + 双层递归库）合并精讲——互不引用、非换代，按「同一问题的两条互补路线」组织；两篇公式链 (1)-(5)/(1)-(4) 逐条白话拆解，正文 4 张论文原图（CC BY 4.0，TeX 源码包矢量 PDF 渲染入 `docs/assets/source-figures/agent-harness/`），配套纯标准库原型 `harness2_lab.py`（selftest + D1–D5 破坏实验 + 字数门禁 count 模式）。 diff --git a/docs/.agents/knowledge-map.md b/docs/.agents/knowledge-map.md index d5113c528..3c2d756a6 100644 --- a/docs/.agents/knowledge-map.md +++ b/docs/.agents/knowledge-map.md @@ -94,6 +94,7 @@ - [精读：claude.dev 全站（177）](../research/agent-harness/177-anthropic-claude-dev-blog.md) — Anthropic 开发者博客全站首读（2026-10-23 快照：16 文 + Mods + 5 视频官方字幕，7 组官方文档补读）：任务账单=轮数×重发×缓存前缀（0.05x 读价与击穿清单）、effort 买彻底不买聪明（Terminal-Bench 3.0 实证）、脚手架生灭律（80% 系统提示删减 / todo 消亡 / RAG→Grep）、七层评审防线（风险 0/1/2 人·agent 批准不算·事故即评测集）与确定性计数棘轮（两周 3.1x·3000+ 变更 0 事故）、无人值守自动化三件套（书签续读/送达确认记账/budget_reached 硬预算）、执行环境三级跳（云会话每任务一 VM→Managed Agents→Claude Tag 按频道记忆·65% PR 口头估计）、mods hook 链三态与 dynamic workflows（ultracode·16 并发）；配套原型 [claude_dev_lab.py](../research/agent-harness/assets/claude_dev_lab.py)（纯标准库确定性，`--selftest` + E1–E4 破坏性实验）与两张 archify 图 - [精读：Can Visual Generation Expand What UMMs Know?（240）](../research/multimodal-models/240-visual-generation-knowledge-transfer.md) — arXiv:2610.21459v1 首读：KTP 知识迁移探针（虚构实体 × 取值只在训练图里 × 纯文字问答 + 重画对照）、直接迁移画得出远多于说得出、对齐集 / 测试集切分下的跨实体迁移与号码 / 书名天花板、交错样本与图像损失消融、先画再读、画对与答对的关联统计；表格复算发现表 2 直接迁移行与对齐行分母可能不同、表 4 六行整数解；配套原型 [ktp_lab.py](../research/multimodal-models/assets/ktp_lab.py)（纯标准库确定性，`--selftest` + E1–E5 破坏性实验）、两张 archify 图与 7 张论文原图（署名学习引用） - [精读：OpenViking 上下文数据库](../research/cognitive-context/014-openviking.md) · [机制映射汇总·认知与上下文 OpenViking](../research/mapping-negentropy.md) — 火山引擎开源 Agent 上下文数据库（[volcengine/OpenViking](https://github.com/volcengine/OpenViking) 钉点 `6227af51`，2026-10-21，v0.5.1.1，AGPL-3.0）的换代重写精读（旧钉点 `14a7b81`，四桶处置 保留9/更新14/删除5/新增7+）：`viking://` 树统一资源/记忆/技能、**User/Peer 归属一等化**（peer 无密钥非租户、写窄读宽、跨 peer 降权 0.1/0.02）、目录级 L0/L1/L2 sidecar 自底向上生成与新鲜度冒泡（256/4000 字符、L0 从 L1 裁出零 LLM、10% 判据，公共 API 不可直写 sidecar）；**检索换代**——上游 #5450（2026-09-29）删除目录递归/父分传播/收敛刹车，现行「全局召回 2×limit 池 → URI 去重 → 单遍精排读 abstract → 阈值 0.1 后置」+ context 六域配额装配（1600 预算、空位补足、降档不截断；`/recall` 弃用）；会话两阶段提交与 schema 身份判据提炼（九类身份字段表、宁重复不错并、WM v0.5.0 默认关、memory_diff 内部有读者对外无回滚）；v0.5 新增 **Gateway 模型代理**（五步一轮、逐字节回放保缓存、默认 13 只读工具/8 写工具须勾选）与治理面（三角色/ACL 默认关/OVE1 信封加密/git 快照前向提交/OVPack v3 明文 ZIP）；证据分级下厂商基准全【三】（LoCoMo 24.20%→82.08% 无消融、竞品列 Mem0 56.62%/SuperMemory 42.99%/百炼 39.55% 自选基线、单轮 RAG 66.87% block(s)" + ] + }, + { + "name": "finite_svg", + "ok": true, + "details": [] + }, + { + "name": "orthogonal_arrows", + "ok": true, + "details": [] + }, + { + "name": "label_route_clearance", + "ok": true, + "details": [] + }, + { + "name": "relationship_crossings", + "ok": true, + "details": [] + }, + { + "name": "relationship_corridors", + "ok": true, + "details": [] + }, + { + "name": "container_border_runs", + "ok": true, + "details": [] + }, + { + "name": "route_rhythm", + "ok": true, + "details": [] + }, + { + "name": "legend_clearance", + "ok": true, + "details": [] + } + ], + "composition": { + "schemaVersion": 1, + "profile": "showcase", + "status": "pass", + "summary": { + "errors": 0, + "warnings": 0 + }, + "metrics": { + "properCrossings": 0, + "resolvedCrossovers": 0, + "ambiguousCorridors": 0, + "arrowheadCollisions": 0, + "containerBorderRuns": 0, + "labelRouteClearanceIssues": 0, + "labelCanvasOverflowIssues": 0, + "minLabelRouteClearance": 64.6, + "desktopReadabilityIssues": 0, + "viewportHeightIssues": 0, + "minProjectedNodeTextPx": 6.348122866894198, + "maxBends": 0, + "routesOverSuggestedBends": 0, + "maxStretch": 1, + "routesOverSuggestedStretch": 0, + "minSegmentPx": 38, + "minInteriorSegmentPx": null, + "shortSegmentCount": 0, + "shortEndpointSegmentCount": 0, + "shortInteriorSegmentCount": 0, + "microSegmentCount": 0 + }, + "suggestedLimits": { + "bendsPerRelationship": 2, + "stretch": 1.35, + "segmentPx": 16, + "microSegmentPx": 8 + }, + "routeReview": { + "crossings": [], + "detours": [] + }, + "leadingSpace": { + "measured": false, + "reviewSuggested": false + }, + "desktopReadability": { + "budgetBasis": "legacy-930", + "readerContract": "declared-wide-v1", + "availableDiagramWidth": 930, + "actualBudgetPx": 930, + "limit": "legacy", + "requestedTargetPx": null, + "requestedTargetMet": null, + "hardFloorPx": 6, + "hardFloorMet": true, + "minimumOwner": { + "kind": "node", + "id": "mem" + }, + "minimumSourceTextPx": 8, + "minimumProjectedTextPx": 6.348122866894198, + "semanticTextCount": 23 + }, + "issues": [] + }, + "provenance": "current", + "deliveryReceiptId": "a57d0505-012e-469e-9d7a-83d105f3cad5" + } + }, + "browser-check": { + "status": "pass", + "exitCode": 0, + "durationMs": 1016, + "command": [ + "/Users/cm.huang/.nvm/versions/node/v24.14.0/bin/node", + "/Users/cm.huang/.agents/skills/archify/bin/archify.mjs", + "browser-check", + "/Users/cm.huang/conductor/workspaces/negentropy/tunis-v4/docs/assets/architecture/cognitive-context/memadapter--three-stage-pipeline.html", + "--json", + "--require-provenance" + ], + "execution": "in-process", + "receipt": { + "schemaVersion": 1, + "ok": true, + "command": "browser-check", + "evidenceKind": "automated-browser", + "status": "pass", + "visualReview": "not-requested", + "provenance": "current", + "deliveryReceiptId": "a57d0505-012e-469e-9d7a-83d105f3cad5", + "artifact": { + "path": "/Users/cm.huang/conductor/workspaces/negentropy/tunis-v4/docs/assets/architecture/cognitive-context/memadapter--three-stage-pipeline.html", + "sha256": "3c00fa4cdb4a80c10d400181186e2ff3fde6cc2f604fb130a1bdb4648252dd39", + "bytes": 756379 + }, + "state": { + "detail": "read", + "motion": "still" + }, + "chrome": { + "status": "available", + "executable": "/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" + }, + "diagnostics": [], + "containment": { + "status": "pass", + "policy": "fit-or-reader-declared-readable-vertical-scroll", + "viewports": [ + { + "width": 1440, + "height": 900, + "theme": "light", + "innerWidth": 1440, + "innerHeight": 900, + "scrollWidth": 1440, + "scrollHeight": 900, + "overflowX": false, + "overflowY": false, + "verticalScrollAccepted": false, + "overflowDisposition": "contained", + "readerLayout": "adaptive", + "readerOverflow": null, + "readerFit": null, + "ok": true, + "readerWidth": 1376, + "diagramWidth": 1346, + "pageComposition": { + "bodyPaddingPx": 12, + "headerPx": 52, + "diagramChromePx": 78, + "svgPx": 464, + "cardsPx": 0, + "viewBoxHeight": 404 + }, + "viewBoxWidth": 1172, + "workflowLanes": [ + { + "frameId": "lane-0", + "heightPx": 119, + "nodeCount": 3, + "nodeIds": [ + "mem", + "s1", + "bc" + ], + "nodeSpanPx": 60, + "spaceAboveNodesPx": 47, + "spaceBelowNodesPx": 13 + }, + { + "frameId": "lane-1", + "heightPx": 119, + "nodeCount": 5, + "nodeIds": [ + "qc", + "s2", + "ui", + "s3", + "ans" + ], + "nodeSpanPx": 60, + "spaceAboveNodesPx": 47, + "spaceBelowNodesPx": 13 + } + ], + "minimumProjectedNodeTextPx": 8, + "minimumProjectedNonEdgeTextPx": 8, + "minimumProjectedEdgeTextPx": 8, + "minimumProjectedNodeText": "不增删不改", + "minimumProjectedNodeTextDetail": "context", + "minimumProjectedNodeTextOwner": { + "kind": "node", + "id": "mem" + }, + "minimumRequiredNodeTextPx": 6, + "readabilityOk": true, + "hasLegend": true, + "hasNavigationDock": true, + "legendDockIntersectionArea": 0, + "dockStageIntersectionArea": 0, + "dockStageGap": 10, + "requiredDockStageGap": 10, + "viewerChromeStageOk": true, + "viewerChromeReserve": 54, + "viewerChromeActive": true, + "viewerChromeOk": true, + "resolvedTheme": "light" + }, + { + "width": 1600, + "height": 1000, + "theme": "light", + "innerWidth": 1600, + "innerHeight": 1000, + "scrollWidth": 1600, + "scrollHeight": 1000, + "overflowX": false, + "overflowY": false, + "verticalScrollAccepted": false, + "overflowDisposition": "contained", + "readerLayout": "adaptive", + "readerOverflow": null, + "readerFit": null, + "ok": true, + "readerWidth": 1536, + "diagramWidth": 1506, + "pageComposition": { + "bodyPaddingPx": 64, + "headerPx": 60, + "diagramChromePx": 84, + "svgPx": 519, + "cardsPx": 0, + "viewBoxHeight": 404 + }, + "viewBoxWidth": 1172, + "workflowLanes": [ + { + "frameId": "lane-0", + "heightPx": 134, + "nodeCount": 3, + "nodeIds": [ + "mem", + "s1", + "bc" + ], + "nodeSpanPx": 67, + "spaceAboveNodesPx": 53, + "spaceBelowNodesPx": 14 + }, + { + "frameId": "lane-1", + "heightPx": 134, + "nodeCount": 5, + "nodeIds": [ + "qc", + "s2", + "ui", + "s3", + "ans" + ], + "nodeSpanPx": 67, + "spaceAboveNodesPx": 53, + "spaceBelowNodesPx": 14 + } + ], + "minimumProjectedNodeTextPx": 8, + "minimumProjectedNonEdgeTextPx": 8, + "minimumProjectedEdgeTextPx": 8, + "minimumProjectedNodeText": "不增删不改", + "minimumProjectedNodeTextDetail": "context", + "minimumProjectedNodeTextOwner": { + "kind": "node", + "id": "mem" + }, + "minimumRequiredNodeTextPx": 6, + "readabilityOk": true, + "hasLegend": true, + "hasNavigationDock": true, + "legendDockIntersectionArea": 0, + "dockStageIntersectionArea": 0, + "dockStageGap": 10, + "requiredDockStageGap": 10, + "viewerChromeStageOk": true, + "viewerChromeReserve": 54, + "viewerChromeActive": true, + "viewerChromeOk": true, + "resolvedTheme": "light" + }, + { + "width": 1920, + "height": 1080, + "theme": "light", + "innerWidth": 1920, + "innerHeight": 1080, + "scrollWidth": 1920, + "scrollHeight": 1080, + "overflowX": false, + "overflowY": false, + "verticalScrollAccepted": false, + "overflowDisposition": "contained", + "readerLayout": "adaptive", + "readerOverflow": null, + "readerFit": null, + "ok": true, + "readerWidth": 1856, + "diagramWidth": 1826, + "pageComposition": { + "bodyPaddingPx": 64, + "headerPx": 60, + "diagramChromePx": 84, + "svgPx": 629, + "cardsPx": 0, + "viewBoxHeight": 404 + }, + "viewBoxWidth": 1172, + "workflowLanes": [ + { + "frameId": "lane-0", + "heightPx": 162, + "nodeCount": 3, + "nodeIds": [ + "mem", + "s1", + "bc" + ], + "nodeSpanPx": 81, + "spaceAboveNodesPx": 64, + "spaceBelowNodesPx": 17 + }, + { + "frameId": "lane-1", + "heightPx": 162, + "nodeCount": 5, + "nodeIds": [ + "qc", + "s2", + "ui", + "s3", + "ans" + ], + "nodeSpanPx": 81, + "spaceAboveNodesPx": 64, + "spaceBelowNodesPx": 17 + } + ], + "minimumProjectedNodeTextPx": 8, + "minimumProjectedNonEdgeTextPx": 8, + "minimumProjectedEdgeTextPx": 8, + "minimumProjectedNodeText": "不增删不改", + "minimumProjectedNodeTextDetail": "context", + "minimumProjectedNodeTextOwner": { + "kind": "node", + "id": "mem" + }, + "minimumRequiredNodeTextPx": 6, + "readabilityOk": true, + "hasLegend": true, + "hasNavigationDock": true, + "legendDockIntersectionArea": 0, + "dockStageIntersectionArea": 0, + "dockStageGap": 10, + "requiredDockStageGap": 10, + "viewerChromeStageOk": true, + "viewerChromeReserve": 54, + "viewerChromeActive": true, + "viewerChromeOk": true, + "resolvedTheme": "light" + }, + { + "width": 2048, + "height": 1320, + "theme": "light", + "innerWidth": 2048, + "innerHeight": 1320, + "scrollWidth": 2048, + "scrollHeight": 1320, + "overflowX": false, + "overflowY": false, + "verticalScrollAccepted": false, + "overflowDisposition": "contained", + "readerLayout": "adaptive", + "readerOverflow": null, + "readerFit": null, + "ok": true, + "readerWidth": 1920, + "diagramWidth": 1870, + "pageComposition": { + "bodyPaddingPx": 64, + "headerPx": 68, + "diagramChromePx": 94, + "svgPx": 645, + "cardsPx": 0, + "viewBoxHeight": 404 + }, + "viewBoxWidth": 1172, + "workflowLanes": [ + { + "frameId": "lane-0", + "heightPx": 166, + "nodeCount": 3, + "nodeIds": [ + "mem", + "s1", + "bc" + ], + "nodeSpanPx": 83, + "spaceAboveNodesPx": 65, + "spaceBelowNodesPx": 18 + }, + { + "frameId": "lane-1", + "heightPx": 166, + "nodeCount": 5, + "nodeIds": [ + "qc", + "s2", + "ui", + "s3", + "ans" + ], + "nodeSpanPx": 83, + "spaceAboveNodesPx": 65, + "spaceBelowNodesPx": 18 + } + ], + "minimumProjectedNodeTextPx": 8, + "minimumProjectedNonEdgeTextPx": 8, + "minimumProjectedEdgeTextPx": 8, + "minimumProjectedNodeText": "不增删不改", + "minimumProjectedNodeTextDetail": "context", + "minimumProjectedNodeTextOwner": { + "kind": "node", + "id": "mem" + }, + "minimumRequiredNodeTextPx": 6, + "readabilityOk": true, + "hasLegend": true, + "hasNavigationDock": true, + "legendDockIntersectionArea": 0, + "dockStageIntersectionArea": 0, + "dockStageGap": 10, + "requiredDockStageGap": 10, + "viewerChromeStageOk": true, + "viewerChromeReserve": 44, + "viewerChromeActive": true, + "viewerChromeOk": true, + "resolvedTheme": "light" + } + ] + }, + "themeStates": { + "status": "pass", + "viewports": [ + { + "width": 1440, + "height": 900, + "requestedTheme": "light", + "resolvedTheme": "light", + "ok": true + }, + { + "width": 1600, + "height": 1000, + "requestedTheme": "light", + "resolvedTheme": "light", + "ok": true + }, + { + "width": 1920, + "height": 1080, + "requestedTheme": "light", + "resolvedTheme": "light", + "ok": true + }, + { + "width": 2048, + "height": 1320, + "requestedTheme": "light", + "resolvedTheme": "light", + "ok": true + }, + { + "width": 1440, + "height": 900, + "requestedTheme": "dark", + "resolvedTheme": "dark", + "ok": true + }, + { + "width": 2048, + "height": 1320, + "requestedTheme": "dark", + "resolvedTheme": "dark", + "ok": true + } + ] + }, + "readability": { + "status": "pass", + "minimumProjectedNodeTextPx": 6, + "viewports": [ + { + "width": 1440, + "height": 900, + "theme": "light", + "innerWidth": 1440, + "innerHeight": 900, + "scrollWidth": 1440, + "scrollHeight": 900, + "overflowX": false, + "overflowY": false, + "verticalScrollAccepted": false, + "overflowDisposition": "contained", + "readerLayout": "adaptive", + "readerOverflow": null, + "readerFit": null, + "ok": true, + "readerWidth": 1376, + "diagramWidth": 1346, + "pageComposition": { + "bodyPaddingPx": 12, + "headerPx": 52, + "diagramChromePx": 78, + "svgPx": 464, + "cardsPx": 0, + "viewBoxHeight": 404 + }, + "viewBoxWidth": 1172, + "workflowLanes": [ + { + "frameId": "lane-0", + "heightPx": 119, + "nodeCount": 3, + "nodeIds": [ + "mem", + "s1", + "bc" + ], + "nodeSpanPx": 60, + "spaceAboveNodesPx": 47, + "spaceBelowNodesPx": 13 + }, + { + "frameId": "lane-1", + "heightPx": 119, + "nodeCount": 5, + "nodeIds": [ + "qc", + "s2", + "ui", + "s3", + "ans" + ], + "nodeSpanPx": 60, + "spaceAboveNodesPx": 47, + "spaceBelowNodesPx": 13 + } + ], + "minimumProjectedNodeTextPx": 8, + "minimumProjectedNonEdgeTextPx": 8, + "minimumProjectedEdgeTextPx": 8, + "minimumProjectedNodeText": "不增删不改", + "minimumProjectedNodeTextDetail": "context", + "minimumProjectedNodeTextOwner": { + "kind": "node", + "id": "mem" + }, + "minimumRequiredNodeTextPx": 6, + "readabilityOk": true, + "hasLegend": true, + "hasNavigationDock": true, + "legendDockIntersectionArea": 0, + "dockStageIntersectionArea": 0, + "dockStageGap": 10, + "requiredDockStageGap": 10, + "viewerChromeStageOk": true, + "viewerChromeReserve": 54, + "viewerChromeActive": true, + "viewerChromeOk": true, + "resolvedTheme": "light" + }, + { + "width": 1600, + "height": 1000, + "theme": "light", + "innerWidth": 1600, + "innerHeight": 1000, + "scrollWidth": 1600, + "scrollHeight": 1000, + "overflowX": false, + "overflowY": false, + "verticalScrollAccepted": false, + "overflowDisposition": "contained", + "readerLayout": "adaptive", + "readerOverflow": null, + "readerFit": null, + "ok": true, + "readerWidth": 1536, + "diagramWidth": 1506, + "pageComposition": { + "bodyPaddingPx": 64, + "headerPx": 60, + "diagramChromePx": 84, + "svgPx": 519, + "cardsPx": 0, + "viewBoxHeight": 404 + }, + "viewBoxWidth": 1172, + "workflowLanes": [ + { + "frameId": "lane-0", + "heightPx": 134, + "nodeCount": 3, + "nodeIds": [ + "mem", + "s1", + "bc" + ], + "nodeSpanPx": 67, + "spaceAboveNodesPx": 53, + "spaceBelowNodesPx": 14 + }, + { + "frameId": "lane-1", + "heightPx": 134, + "nodeCount": 5, + "nodeIds": [ + "qc", + "s2", + "ui", + "s3", + "ans" + ], + "nodeSpanPx": 67, + "spaceAboveNodesPx": 53, + "spaceBelowNodesPx": 14 + } + ], + "minimumProjectedNodeTextPx": 8, + "minimumProjectedNonEdgeTextPx": 8, + "minimumProjectedEdgeTextPx": 8, + "minimumProjectedNodeText": "不增删不改", + "minimumProjectedNodeTextDetail": "context", + "minimumProjectedNodeTextOwner": { + "kind": "node", + "id": "mem" + }, + "minimumRequiredNodeTextPx": 6, + "readabilityOk": true, + "hasLegend": true, + "hasNavigationDock": true, + "legendDockIntersectionArea": 0, + "dockStageIntersectionArea": 0, + "dockStageGap": 10, + "requiredDockStageGap": 10, + "viewerChromeStageOk": true, + "viewerChromeReserve": 54, + "viewerChromeActive": true, + "viewerChromeOk": true, + "resolvedTheme": "light" + }, + { + "width": 1920, + "height": 1080, + "theme": "light", + "innerWidth": 1920, + "innerHeight": 1080, + "scrollWidth": 1920, + "scrollHeight": 1080, + "overflowX": false, + "overflowY": false, + "verticalScrollAccepted": false, + "overflowDisposition": "contained", + "readerLayout": "adaptive", + "readerOverflow": null, + "readerFit": null, + "ok": true, + "readerWidth": 1856, + "diagramWidth": 1826, + "pageComposition": { + "bodyPaddingPx": 64, + "headerPx": 60, + "diagramChromePx": 84, + "svgPx": 629, + "cardsPx": 0, + "viewBoxHeight": 404 + }, + "viewBoxWidth": 1172, + "workflowLanes": [ + { + "frameId": "lane-0", + "heightPx": 162, + "nodeCount": 3, + "nodeIds": [ + "mem", + "s1", + "bc" + ], + "nodeSpanPx": 81, + "spaceAboveNodesPx": 64, + "spaceBelowNodesPx": 17 + }, + { + "frameId": "lane-1", + "heightPx": 162, + "nodeCount": 5, + "nodeIds": [ + "qc", + "s2", + "ui", + "s3", + "ans" + ], + "nodeSpanPx": 81, + "spaceAboveNodesPx": 64, + "spaceBelowNodesPx": 17 + } + ], + "minimumProjectedNodeTextPx": 8, + "minimumProjectedNonEdgeTextPx": 8, + "minimumProjectedEdgeTextPx": 8, + "minimumProjectedNodeText": "不增删不改", + "minimumProjectedNodeTextDetail": "context", + "minimumProjectedNodeTextOwner": { + "kind": "node", + "id": "mem" + }, + "minimumRequiredNodeTextPx": 6, + "readabilityOk": true, + "hasLegend": true, + "hasNavigationDock": true, + "legendDockIntersectionArea": 0, + "dockStageIntersectionArea": 0, + "dockStageGap": 10, + "requiredDockStageGap": 10, + "viewerChromeStageOk": true, + "viewerChromeReserve": 54, + "viewerChromeActive": true, + "viewerChromeOk": true, + "resolvedTheme": "light" + }, + { + "width": 2048, + "height": 1320, + "theme": "light", + "innerWidth": 2048, + "innerHeight": 1320, + "scrollWidth": 2048, + "scrollHeight": 1320, + "overflowX": false, + "overflowY": false, + "verticalScrollAccepted": false, + "overflowDisposition": "contained", + "readerLayout": "adaptive", + "readerOverflow": null, + "readerFit": null, + "ok": true, + "readerWidth": 1920, + "diagramWidth": 1870, + "pageComposition": { + "bodyPaddingPx": 64, + "headerPx": 68, + "diagramChromePx": 94, + "svgPx": 645, + "cardsPx": 0, + "viewBoxHeight": 404 + }, + "viewBoxWidth": 1172, + "workflowLanes": [ + { + "frameId": "lane-0", + "heightPx": 166, + "nodeCount": 3, + "nodeIds": [ + "mem", + "s1", + "bc" + ], + "nodeSpanPx": 83, + "spaceAboveNodesPx": 65, + "spaceBelowNodesPx": 18 + }, + { + "frameId": "lane-1", + "heightPx": 166, + "nodeCount": 5, + "nodeIds": [ + "qc", + "s2", + "ui", + "s3", + "ans" + ], + "nodeSpanPx": 83, + "spaceAboveNodesPx": 65, + "spaceBelowNodesPx": 18 + } + ], + "minimumProjectedNodeTextPx": 8, + "minimumProjectedNonEdgeTextPx": 8, + "minimumProjectedEdgeTextPx": 8, + "minimumProjectedNodeText": "不增删不改", + "minimumProjectedNodeTextDetail": "context", + "minimumProjectedNodeTextOwner": { + "kind": "node", + "id": "mem" + }, + "minimumRequiredNodeTextPx": 6, + "readabilityOk": true, + "hasLegend": true, + "hasNavigationDock": true, + "legendDockIntersectionArea": 0, + "dockStageIntersectionArea": 0, + "dockStageGap": 10, + "requiredDockStageGap": 10, + "viewerChromeStageOk": true, + "viewerChromeReserve": 44, + "viewerChromeActive": true, + "viewerChromeOk": true, + "resolvedTheme": "light" + } + ] + }, + "viewerChrome": { + "status": "pass", + "viewports": [ + { + "width": 1440, + "height": 900, + "theme": "light", + "innerWidth": 1440, + "innerHeight": 900, + "scrollWidth": 1440, + "scrollHeight": 900, + "overflowX": false, + "overflowY": false, + "verticalScrollAccepted": false, + "overflowDisposition": "contained", + "readerLayout": "adaptive", + "readerOverflow": null, + "readerFit": null, + "ok": true, + "readerWidth": 1376, + "diagramWidth": 1346, + "pageComposition": { + "bodyPaddingPx": 12, + "headerPx": 52, + "diagramChromePx": 78, + "svgPx": 464, + "cardsPx": 0, + "viewBoxHeight": 404 + }, + "viewBoxWidth": 1172, + "workflowLanes": [ + { + "frameId": "lane-0", + "heightPx": 119, + "nodeCount": 3, + "nodeIds": [ + "mem", + "s1", + "bc" + ], + "nodeSpanPx": 60, + "spaceAboveNodesPx": 47, + "spaceBelowNodesPx": 13 + }, + { + "frameId": "lane-1", + "heightPx": 119, + "nodeCount": 5, + "nodeIds": [ + "qc", + "s2", + "ui", + "s3", + "ans" + ], + "nodeSpanPx": 60, + "spaceAboveNodesPx": 47, + "spaceBelowNodesPx": 13 + } + ], + "minimumProjectedNodeTextPx": 8, + "minimumProjectedNonEdgeTextPx": 8, + "minimumProjectedEdgeTextPx": 8, + "minimumProjectedNodeText": "不增删不改", + "minimumProjectedNodeTextDetail": "context", + "minimumProjectedNodeTextOwner": { + "kind": "node", + "id": "mem" + }, + "minimumRequiredNodeTextPx": 6, + "readabilityOk": true, + "hasLegend": true, + "hasNavigationDock": true, + "legendDockIntersectionArea": 0, + "dockStageIntersectionArea": 0, + "dockStageGap": 10, + "requiredDockStageGap": 10, + "viewerChromeStageOk": true, + "viewerChromeReserve": 54, + "viewerChromeActive": true, + "viewerChromeOk": true, + "resolvedTheme": "light" + }, + { + "width": 1600, + "height": 1000, + "theme": "light", + "innerWidth": 1600, + "innerHeight": 1000, + "scrollWidth": 1600, + "scrollHeight": 1000, + "overflowX": false, + "overflowY": false, + "verticalScrollAccepted": false, + "overflowDisposition": "contained", + "readerLayout": "adaptive", + "readerOverflow": null, + "readerFit": null, + "ok": true, + "readerWidth": 1536, + "diagramWidth": 1506, + "pageComposition": { + "bodyPaddingPx": 64, + "headerPx": 60, + "diagramChromePx": 84, + "svgPx": 519, + "cardsPx": 0, + "viewBoxHeight": 404 + }, + "viewBoxWidth": 1172, + "workflowLanes": [ + { + "frameId": "lane-0", + "heightPx": 134, + "nodeCount": 3, + "nodeIds": [ + "mem", + "s1", + "bc" + ], + "nodeSpanPx": 67, + "spaceAboveNodesPx": 53, + "spaceBelowNodesPx": 14 + }, + { + "frameId": "lane-1", + "heightPx": 134, + "nodeCount": 5, + "nodeIds": [ + "qc", + "s2", + "ui", + "s3", + "ans" + ], + "nodeSpanPx": 67, + "spaceAboveNodesPx": 53, + "spaceBelowNodesPx": 14 + } + ], + "minimumProjectedNodeTextPx": 8, + "minimumProjectedNonEdgeTextPx": 8, + "minimumProjectedEdgeTextPx": 8, + "minimumProjectedNodeText": "不增删不改", + "minimumProjectedNodeTextDetail": "context", + "minimumProjectedNodeTextOwner": { + "kind": "node", + "id": "mem" + }, + "minimumRequiredNodeTextPx": 6, + "readabilityOk": true, + "hasLegend": true, + "hasNavigationDock": true, + "legendDockIntersectionArea": 0, + "dockStageIntersectionArea": 0, + "dockStageGap": 10, + "requiredDockStageGap": 10, + "viewerChromeStageOk": true, + "viewerChromeReserve": 54, + "viewerChromeActive": true, + "viewerChromeOk": true, + "resolvedTheme": "light" + }, + { + "width": 1920, + "height": 1080, + "theme": "light", + "innerWidth": 1920, + "innerHeight": 1080, + "scrollWidth": 1920, + "scrollHeight": 1080, + "overflowX": false, + "overflowY": false, + "verticalScrollAccepted": false, + "overflowDisposition": "contained", + "readerLayout": "adaptive", + "readerOverflow": null, + "readerFit": null, + "ok": true, + "readerWidth": 1856, + "diagramWidth": 1826, + "pageComposition": { + "bodyPaddingPx": 64, + "headerPx": 60, + "diagramChromePx": 84, + "svgPx": 629, + "cardsPx": 0, + "viewBoxHeight": 404 + }, + "viewBoxWidth": 1172, + "workflowLanes": [ + { + "frameId": "lane-0", + "heightPx": 162, + "nodeCount": 3, + "nodeIds": [ + "mem", + "s1", + "bc" + ], + "nodeSpanPx": 81, + "spaceAboveNodesPx": 64, + "spaceBelowNodesPx": 17 + }, + { + "frameId": "lane-1", + "heightPx": 162, + "nodeCount": 5, + "nodeIds": [ + "qc", + "s2", + "ui", + "s3", + "ans" + ], + "nodeSpanPx": 81, + "spaceAboveNodesPx": 64, + "spaceBelowNodesPx": 17 + } + ], + "minimumProjectedNodeTextPx": 8, + "minimumProjectedNonEdgeTextPx": 8, + "minimumProjectedEdgeTextPx": 8, + "minimumProjectedNodeText": "不增删不改", + "minimumProjectedNodeTextDetail": "context", + "minimumProjectedNodeTextOwner": { + "kind": "node", + "id": "mem" + }, + "minimumRequiredNodeTextPx": 6, + "readabilityOk": true, + "hasLegend": true, + "hasNavigationDock": true, + "legendDockIntersectionArea": 0, + "dockStageIntersectionArea": 0, + "dockStageGap": 10, + "requiredDockStageGap": 10, + "viewerChromeStageOk": true, + "viewerChromeReserve": 54, + "viewerChromeActive": true, + "viewerChromeOk": true, + "resolvedTheme": "light" + }, + { + "width": 2048, + "height": 1320, + "theme": "light", + "innerWidth": 2048, + "innerHeight": 1320, + "scrollWidth": 2048, + "scrollHeight": 1320, + "overflowX": false, + "overflowY": false, + "verticalScrollAccepted": false, + "overflowDisposition": "contained", + "readerLayout": "adaptive", + "readerOverflow": null, + "readerFit": null, + "ok": true, + "readerWidth": 1920, + "diagramWidth": 1870, + "pageComposition": { + "bodyPaddingPx": 64, + "headerPx": 68, + "diagramChromePx": 94, + "svgPx": 645, + "cardsPx": 0, + "viewBoxHeight": 404 + }, + "viewBoxWidth": 1172, + "workflowLanes": [ + { + "frameId": "lane-0", + "heightPx": 166, + "nodeCount": 3, + "nodeIds": [ + "mem", + "s1", + "bc" + ], + "nodeSpanPx": 83, + "spaceAboveNodesPx": 65, + "spaceBelowNodesPx": 18 + }, + { + "frameId": "lane-1", + "heightPx": 166, + "nodeCount": 5, + "nodeIds": [ + "qc", + "s2", + "ui", + "s3", + "ans" + ], + "nodeSpanPx": 83, + "spaceAboveNodesPx": 65, + "spaceBelowNodesPx": 18 + } + ], + "minimumProjectedNodeTextPx": 8, + "minimumProjectedNonEdgeTextPx": 8, + "minimumProjectedEdgeTextPx": 8, + "minimumProjectedNodeText": "不增删不改", + "minimumProjectedNodeTextDetail": "context", + "minimumProjectedNodeTextOwner": { + "kind": "node", + "id": "mem" + }, + "minimumRequiredNodeTextPx": 6, + "readabilityOk": true, + "hasLegend": true, + "hasNavigationDock": true, + "legendDockIntersectionArea": 0, + "dockStageIntersectionArea": 0, + "dockStageGap": 10, + "requiredDockStageGap": 10, + "viewerChromeStageOk": true, + "viewerChromeReserve": 44, + "viewerChromeActive": true, + "viewerChromeOk": true, + "resolvedTheme": "light" + } + ] + }, + "captures": { + "status": "not-requested", + "screenshots": [], + "contactSheet": null + }, + "sidecars": { + "receipt": "memadapter--three-stage-pipeline.browser-check.json", + "files": [] + } + } + } + }, + "diagnostics": [], + "evidence": { + "receipt": "/Users/cm.huang/conductor/workspaces/negentropy/tunis-v4/docs/assets/architecture/cognitive-context/memadapter--three-stage-pipeline.finalize.json", + "summaryReceipt": "/Users/cm.huang/conductor/workspaces/negentropy/tunis-v4/docs/assets/architecture/cognitive-context/memadapter--three-stage-pipeline.finalize-summary.json", + "browserCheckReceipt": "/Users/cm.huang/conductor/workspaces/negentropy/tunis-v4/docs/assets/architecture/cognitive-context/memadapter--three-stage-pipeline.browser-check.json" + }, + "visualReview": "not-requested", + "finishedAt": "2026-10-28T04:02:11.558Z", + "durationMs": 1874 +} diff --git a/docs/assets/architecture/cognitive-context/memadapter--three-stage-pipeline.html b/docs/assets/architecture/cognitive-context/memadapter--three-stage-pipeline.html new file mode 100644 index 000000000..a9b52a2aa --- /dev/null +++ b/docs/assets/architecture/cognitive-context/memadapter--three-stage-pipeline.html @@ -0,0 +1,13741 @@ + + + + + + + + MemAdapter 检索后三阶段：从记忆边界到证据锚定 + + + + + + + + +
+ +
+
+
+

MemAdapter 检索后三阶段：从记忆边界到证据锚定

+
+
+ + + + +
+ + MemAdapter 检索后三阶段：从记忆边界到证据锚定 + 由 Archify 生成的工作流图。 + + + + + + + + + + + + + + + + + + + + + + + + + 01 / 任务无关 · 不看当前查询 + + + 02 / 任务条件化 · 对照当前任务 + + + + + + + + + + + + + + + + + + + 检索记忆集 ℳ · 不增删不改 · 任务无关 · 不看当前查询 + + + + 检索记忆集 ℳ + 不增删不改 + + + + ① 反事实归纳 · 归纳条件边界（式1–3） · 任务无关 · 不看当前查询 + + + + ① 反事实归纳 + 归纳条件边界（式1–3） + + + + 条件边界卡 ℬ · 任务无关 · 何时可用 · 任务无关 · 不看当前查询 + + + + 条件边界卡 ℬ + 任务无关 · 何时可用 + + + + 当前请求与证据 q, c · 任务目标 · 证据 · 任务条件化 · 对照当前任务 + + + + 当前请求与证据 q, c + 任务目标 · 证据 + + + + ② 上下文感知反思 · 任务校准（式4） · 任务条件化 · 对照当前任务 + + + + ② 上下文感知反思 + 任务校准（式4） + + + + 使用指令 𝒰 · 能干什么 · 到哪为止 · 任务条件化 · 对照当前任务 + + + + 使用指令 𝒰 + 能干什么 · 到哪为止 + + + + ③ 证据锚定推理 · 证据锚定＋核验（式5） · 任务条件化 · 对照当前任务 + + + + ③ 证据锚定推理 + 证据锚定＋核验（式5） + + + + 最终答案 y · ℒ 内部核验不外露 · 任务条件化 · 对照当前任务 + + + + 最终答案 y + ℒ 内部核验不外露 + + + + + + 边界输入 + + + + 整批记忆（不含查询） + + + + 当前任务与证据 + + + + 归纳 + + + + 译成指令 + + + + 核验通过后交卷 + + + + 指令约束 + + + + + 图例 + + + Agent 逻辑 + + + + 上下文 / 追踪 + + + + 外部系统 + + + +

+ + + + + + + + + +
+ + + +
+ + + + diff --git a/docs/assets/mermaid/README.md b/docs/assets/mermaid/README.md index 30c59f99f..19b96bd26 100644 --- a/docs/assets/mermaid/README.md +++ b/docs/assets/mermaid/README.md @@ -8,7 +8,7 @@ > > **范围**：按**文档**而非目录界定。已纳入：现役系统文档（`docs/concepts/`、`docs/reference/{perceives,wiki}/`、根与 i18n README、 > `apps/` README、`docs/.agents/` 巡检文档），以及已走完本管线的研究文献——`docs/research/cognitive-context/` 的 -> Horizon Context 精读、Context Layer 蓝图与 OpenViking 精读（下表 `cognitive-context/` 分节）、 +> Horizon Context 精读、Context Layer 蓝图、OpenViking 精读与 Memadapter 精读（下表 `cognitive-context/` 分节）、 > `docs/research/agent-harness/` 的五层 Harness 精读、并发独立重学（176）、AI Native 手册精读（180）与 Hermes Agent 精读（190，下表 `agent-harness/` 分节）、 > `docs/research/self-evolution/` 的 Dream-RSI 精读（下表 `self-evolution/` 分节）、 > `docs/research/agent-infra/` 的 Agent Skills 规范精读（230）与 Jev（System One 决策模型）精读（下表 `agent-infra/` 分节）、 @@ -281,6 +281,7 @@ | slug | 源文档锚点 | 类型 | 产物 | 状态 | 备注 | | :--- | :--- | :--- | :--- | :--- | :--- | +| [memadapter--three-stage-pipeline](./cognitive-context/memadapter--three-stage-pipeline.mmd) | cognitive-context/015-memadapter-counterfactual-adaptation.md §3 机制（arXiv:2610.05162v1 §3） | workflow | ✓ | done | 检索后三阶段主链（记忆→反事实归纳→边界卡→反思→使用指令→证据锚定→答案）；双泳道区分「任务无关·不看当前查询」与「任务条件化·对照当前任务」；三卡：为何不看查询/边界≠过滤/原型实测阶梯 | | [context-layer--failure-map](./cognitive-context/context-layer--failure-map.mmd) | cognitive-context/013-context-layer-blueprint.md §1 失效模式登记簿 | architecture | ✓ | done | 十簇蛇形主链×三条问题带；sublabel=代表失效+行数、tag=主要防护；三卡：十簇×五轴对位/绕开测试/回路在线三条件 | | [context-layer--five-sources](./cognitive-context/context-layer--five-sources.mmd) | 已退役（013 无正文引用——五源为资产轴子集投影；v4 重制后 pipeline.toml 无引用，资产仅存档） | architecture | ✓ | retired | 退役：投影关系见 013 §4 五源投影注记 | | [context-layer-blueprint--industry-landscape](./cognitive-context/context-layer-blueprint--industry-landscape.mmd) | cognitive-context/013-context-layer-blueprint.md §11.1 描述还是执行（2026-10-10 微改重锚） | architecture | ✓ | done | 微改：四路线谱系+Palantir 异类收敛「治理往 SQL 生成之前挪」；+Forrester 双阵营与传输vs含义判词、Gartner 60% projection 带口径注 | diff --git a/docs/assets/mermaid/cognitive-context/memadapter--three-stage-pipeline.mmd b/docs/assets/mermaid/cognitive-context/memadapter--three-stage-pipeline.mmd new file mode 100644 index 000000000..a76693db6 --- /dev/null +++ b/docs/assets/mermaid/cognitive-context/memadapter--three-stage-pipeline.mmd @@ -0,0 +1,17 @@ +%%source: docs/research/cognitive-context/015-memadapter-counterfactual-adaptation.md §3 机制（arXiv:2610.05162v1 §3，MemAdapter 三组件） +%%slug: cognitive-context/memadapter--three-stage-pipeline +%%type: workflow +%%derived: archify workflow v2（检索后三阶段：反事实归纳边界 → 任务条件化指令 → 证据锚定生成；双泳道区分「不看当前查询」与「对照当前任务」两段） +flowchart LR + MEM["检索记忆集 ℳ
记忆系统已给出的候选 · 不增删不改"] + S1["① 反事实归纳
枚举假想任务→评估恰当贡献→归并条件边界（式1–3）"] + BC["条件边界卡 ℬ
任务无关的使用边界"] + S2["② 上下文感知反思
对照当前任务联合校准（式4）"] + UI["使用指令 𝒰
支持什么·影响哪部分·到哪为止"] + S3["③ 证据锚定推理
证据为锚生成 y＋支撑轨迹 ℒ 逐句核验（式5）"] + ANS["最终答案 y
ℒ 只做内部核验不外露"] + QC["当前请求与证据 q, c
任务目标·可用证据与约束"] + MEM -->|"整批记忆（不含查询）"| S1 -->|"归纳"| BC + BC -->|"边界输入"| S2 + QC -->|"当前任务与证据"| S2 + S2 -->|"译成指令"| UI -->|"指令约束"| S3 -->|"核验通过后交卷"| ANS diff --git a/docs/assets/source-figures/README.md b/docs/assets/source-figures/README.md index 0030fd97e..ddcb120fc 100644 --- a/docs/assets/source-figures/README.md +++ b/docs/assets/source-figures/README.md @@ -51,6 +51,17 @@ KTP 论文图 1–7 由 TeX 源码包矢量 PDF 渲染为 PNG，2026-10-26 自 ` | 9 | [openviking--src-09-vikingmem-schema.png](./cognitive-context/openviking--src-09-vikingmem-schema.png) | [VikingMem, arXiv:2605.29640v3](https://arxiv.org/html/2605.29640v3#S2.F1) Figure 1（J. Fu *et al.*） | CC BY-NC-SA 4.0 | 原样 | | 10 | [openviking--src-10-product-retrieval-claims.png](./cognitive-context/openviking--src-10-product-retrieval-claims.png) | [openviking.ai/product](https://openviking.ai/product)「三层结构，按需加载」与「理解结构的检索」两节（前端 bundle `index-BfoH6ZUa.js`） | 版权归原作者，仅作学习引用 | 无头浏览器暗色 2× 截图，隐藏吸顶导航 | +### cognitive-context · Memadapter 论文精读（[015](../../research/cognitive-context/015-memadapter-counterfactual-adaptation.md)） + +Memadapter 论文图 1–3 取自 [arXiv:2610.05162v1](https://arxiv.org/html/2610.05162v1) HTML 版（抓取 2026-10-28；论文摘要页许可 CC BY 4.0，允许署名再分发），编号沿用正文「原文图 N」。 + +| 文件 | 出处 | 许可 | 处理 | +| --- | --- | --- | --- | +| [memadapter--fig1-two-paradigms.png](./cognitive-context/memadapter--fig1-two-paradigms.png) | [arXiv:2610.05162v1](https://arxiv.org/html/2610.05162v1#S1.F1) 原文图 1（R. Ning *et al.*） | CC BY 4.0 | 原样 | +| [memadapter--fig2a-accuracy-with-memory.png](./cognitive-context/memadapter--fig2a-accuracy-with-memory.png) | [arXiv:2610.05162v1](https://arxiv.org/html/2610.05162v1#S1.F2) 原文图 2(a)（R. Ning *et al.*） | CC BY 4.0 | SVG 无头浏览器 2× 栅格化（白底，595×346 视口） | +| [memadapter--fig2b-cardiologist-case.png](./cognitive-context/memadapter--fig2b-cardiologist-case.png) | [arXiv:2610.05162v1](https://arxiv.org/html/2610.05162v1#S1.F2) 原文图 2(b)（R. Ning *et al.*） | CC BY 4.0 | 原样 | +| [memadapter--fig3-pipeline.png](./cognitive-context/memadapter--fig3-pipeline.png) | [arXiv:2610.05162v1](https://arxiv.org/html/2610.05162v1#S2.F3) 原文图 3（R. Ning *et al.*） | CC BY 4.0 | 原样 | + ### agent-harness · Harness 到 Harness² 精读（[179](../../research/agent-harness/179-harness-to-harness2.md)） 抓取基准为 arXiv:2610.14224v1 TeX 源码包（e-print，2026-10-28 下载），矢量 PDF 经 pdftoppm 200dpi 渲染 PNG（RGB 白底，无透明通道）；论文许可 CC BY 4.0。 diff --git a/docs/assets/source-figures/cognitive-context/memadapter--fig1-two-paradigms.png b/docs/assets/source-figures/cognitive-context/memadapter--fig1-two-paradigms.png new file mode 100644 index 000000000..bc2434a16 Binary files /dev/null and b/docs/assets/source-figures/cognitive-context/memadapter--fig1-two-paradigms.png differ diff --git a/docs/assets/source-figures/cognitive-context/memadapter--fig2a-accuracy-with-memory.png b/docs/assets/source-figures/cognitive-context/memadapter--fig2a-accuracy-with-memory.png new file mode 100644 index 000000000..51e013e2d Binary files /dev/null and b/docs/assets/source-figures/cognitive-context/memadapter--fig2a-accuracy-with-memory.png differ diff --git a/docs/assets/source-figures/cognitive-context/memadapter--fig2b-cardiologist-case.png b/docs/assets/source-figures/cognitive-context/memadapter--fig2b-cardiologist-case.png new file mode 100644 index 000000000..332885376 Binary files /dev/null and b/docs/assets/source-figures/cognitive-context/memadapter--fig2b-cardiologist-case.png differ diff --git a/docs/assets/source-figures/cognitive-context/memadapter--fig3-pipeline.png b/docs/assets/source-figures/cognitive-context/memadapter--fig3-pipeline.png new file mode 100644 index 000000000..3b5ad2978 Binary files /dev/null and b/docs/assets/source-figures/cognitive-context/memadapter--fig3-pipeline.png differ diff --git a/docs/diagrams/memadapter--three-stage-pipeline/memadapter--three-stage-pipeline.json b/docs/diagrams/memadapter--three-stage-pipeline/memadapter--three-stage-pipeline.json new file mode 100644 index 000000000..e5e0b7881 --- /dev/null +++ b/docs/diagrams/memadapter--three-stage-pipeline/memadapter--three-stage-pipeline.json @@ -0,0 +1,181 @@ +{ + "schema_version": 2, + "diagram_type": "workflow", + "meta": { + "title": "MemAdapter 检索后三阶段：从记忆边界到证据锚定", + "locale": "zh-CN", + "quality_profile": "showcase", + "output": "docs/assets/architecture/cognitive-context/memadapter--three-stage-pipeline.html" + }, + "lanes": [ + { + "id": "task_agnostic", + "label": "任务无关 · 不看当前查询" + }, + { + "id": "task_tuned", + "label": "任务条件化 · 对照当前任务" + } + ], + "mainPath": [ + "mem", + "s1", + "bc", + "s2", + "ui", + "s3", + "ans" + ], + "nodes": [ + { + "id": "mem", + "lane": "task_agnostic", + "col": 0, + "type": "database", + "label": "检索记忆集 ℳ", + "sublabel": "不增删不改", + "icon": "database", + "width": 100 + }, + { + "id": "s1", + "lane": "task_agnostic", + "col": 1, + "type": "backend", + "label": "① 反事实归纳", + "sublabel": "归纳条件边界（式1–3）", + "icon": "active", + "width": 135 + }, + { + "id": "bc", + "lane": "task_agnostic", + "col": 2, + "type": "database", + "label": "条件边界卡 ℬ", + "sublabel": "任务无关 · 何时可用", + "icon": "database", + "width": 125 + }, + { + "id": "qc", + "lane": "task_tuned", + "col": 0, + "type": "external", + "label": "当前请求与证据 q, c", + "sublabel": "任务目标 · 证据", + "icon": "start", + "width": 140 + }, + { + "id": "s2", + "lane": "task_tuned", + "col": 2, + "type": "backend", + "label": "② 上下文感知反思", + "sublabel": "任务校准（式4）", + "icon": "active", + "width": 125 + }, + { + "id": "ui", + "lane": "task_tuned", + "col": 3, + "type": "database", + "label": "使用指令 𝒰", + "sublabel": "能干什么 · 到哪为止", + "icon": "database", + "width": 120 + }, + { + "id": "s3", + "lane": "task_tuned", + "col": 4, + "type": "backend", + "label": "③ 证据锚定推理", + "sublabel": "证据锚定＋核验（式5）", + "icon": "active", + "width": 135 + }, + { + "id": "ans", + "lane": "task_tuned", + "col": 5, + "type": "external", + "label": "最终答案 y", + "sublabel": "ℒ 内部核验不外露", + "icon": "success", + "width": 115 + } + ], + "edges": [ + { + "id": "mem_s1", + "from": "mem", + "to": "s1", + "label": "整批记忆（不含查询）" + }, + { + "id": "s1_bc", + "from": "s1", + "to": "bc", + "label": "归纳" + }, + { + "id": "bc_s2", + "from": "bc", + "to": "s2", + "label": "边界输入" + }, + { + "id": "qc_s2", + "from": "qc", + "to": "s2", + "label": "当前任务与证据" + }, + { + "id": "s2_ui", + "from": "s2", + "to": "ui", + "label": "译成指令" + }, + { + "id": "ui_s3", + "from": "ui", + "to": "s3", + "label": "指令约束" + }, + { + "id": "s3_ans", + "from": "s3", + "to": "ans", + "label": "核验通过后交卷" + } + ], + "cards": [ + { + "dot": "violet", + "title": "① 为什么不看当前查询", + "items": [ + "边界要刻画记忆自身的完整风险画像：先看任务只会暴露当前一种任务下的表现，换任务风险照样冒出来", + "任务无关归纳的边界可跨任务复用；论文消融实测：单这一步即贡献过半总增益（NaiveRAG 70.25→84.26）" + ] + }, + { + "dot": "emerald", + "title": "边界 ≠ 过滤", + "items": [ + "记忆全部保留，治理的是每条的影响范围——支持什么、影响哪部分、不能证成什么", + "删「坏记忆」救不了「准确记忆的越界使用」：初步研究只用准确相关记忆，整体准确率仍 96.0%→77.3%" + ] + }, + { + "dot": "amber", + "title": "玩具原型实测阶梯", + "items": [ + "直接生成 3/9 → +边界卡 7/9 → +任务校准 8/9 → +逐句核验 9/9", + "拆任一阶段都有具体坏法：无边界时「相关即放行」，条件性陷阱全数复发" + ] + } + ] +} diff --git a/docs/research/cognitive-context/015-memadapter-counterfactual-adaptation.md b/docs/research/cognitive-context/015-memadapter-counterfactual-adaptation.md new file mode 100644 index 000000000..cc39437e8 --- /dev/null +++ b/docs/research/cognitive-context/015-memadapter-counterfactual-adaptation.md @@ -0,0 +1,163 @@ +--- +sidebar_position: 8 +title: "Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy 精读与通俗拆解（MemAdapter 记忆谄媚反事实适配）" +description: "arXiv:2610.05162v1 精读（2026-10-04，CC BY 4.0，吉林大学+厦门大学）：带长期记忆的 agent 会顺着用户旧观点答话，且只用准确相关的记忆也救不了——受控实验 96.0%→77.3%、翻错全部跟随记忆；MemAdapter 不动记忆库与检索，在检索后加三道纯提示工序：反事实归纳任务无关使用边界（式1–3）→ 对照当前任务译成使用指令（式4）→ 证据锚定生成＋内部支撑轨迹核验（式5）；三基准×五记忆系统×三骨干验证，PersistBench 谄媚失败率 FR@3 最多降 31pp（附录更深口径 −35.5，DeepSeek 单骨干）、五系统平均分极差从 34.68 收敛到 4.97（主表与附录表 7 在有益使用指标上互相矛盾）；配 4 张论文原图、1 张 archify 三阶段图与纯标准库九场景原型（变体阶梯 3→7→8→9、四次破坏实验）。" +--- + +# Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy 精读与通俗拆解 + +> [1] "MemAdapter: Counterfactual adaptation against memory-induced sycophancy," arXiv:2610.05162v1, Oct. 4, 2026. [Online]. Available: https://arxiv.org/html/2610.05162v1 （访问 2026-10-28；唯一版本无勘误；代码已开源 [1]；作者见参考）。许可 CC BY 4.0。 + +> [!NOTE] 精读范围 +> 论文与附录 A–H 全量（三图五式、全部结果表、算法与提示模板）；配套原型 `assets/memadapter_lab.py` 九场景实测。无独立第三方复现可引，见 §7。 + +**一句话定位**：MemAdapter 不决定「哪些记忆可用」，而决定「每条记忆在当前任务里能干什么、不能干什么」——把记忆安全从内容属性改写成任务条件属性。 + +> [!TIP] 白话主线 +> 带长期记忆的助手会顺着用户历史观点说话，哪怕证据说另一回事。难点：这不是记忆存错——只用百分百准确且相关的记忆，准确率照样从 96.0% 掉到 77.3%，翻错全部跟着记忆走；风险不在记忆内容里，而在记忆被放进什么任务里用。MemAdapter 在检索后加三道工序：先不看当前问题，给每条记忆枚举「在哪些任务里能支持什么、不能拿来证明什么」；再对照当前任务把边界译成使用指令；最后按指令以证据为锚生成、内部核验后交卷。五记忆系统（PersistBench）上，谄媚失败率（FR@3）最多降 31 个百分点。 + +## 1. 它要解决什么问题 + +「记忆诱发谄媚」（memory-induced sycophancy）：持久记忆把用户的旧信念、偏好、旧判断反复塞回后续推理，agent 于是过度对齐用户的历史观点——即使这些观点不准确、过时或与证据矛盾 [1]。长程任务里最凶：记忆活得比形成它的对话更久，一时的影响固化成行为偏差。 + +论文先用受控实验拆掉「记忆错了才出事」的直觉 [1]：三个评测基准各抽 25 个实例（共 75 个，人工逐条审核），保证每条记忆**准确、良性、相关、且不足以单独定答案**；同一问题有/无记忆各答一遍。结果：整体准确率 96.0%→77.3%，答对样本里 19.4% 翻错，且**每一次翻错都朝记忆对齐的方向**——掉分由记忆驱动，不是「上下文变长」的副作用。 + +![原文图 2(a)：只用准确相关记忆，有无记忆的准确率对比](../../assets/source-figures/cognitive-context/memadapter--fig2a-accuracy-with-memory.png) + +来源：Ning et al. [1] 原文图 2(a)，CC BY 4.0。 + +另一半证据：同一记忆跨任务的变化 [1]——「用户是心脏病学家」，讲新指南时正当——可以用专业术语、讲临床区别；替病人选方案时越界——凭身份压过当前病历证据就是谄媚。同一记忆，两个任务，两种正当影响。。 + +![原文图 2(b)：同一记忆跨上下文应有不同影响的心脏病学家案例](../../assets/source-figures/cognitive-context/memadapter--fig2b-cardiologist-case.png) + +来源：Ning et al. [1] 原文图 2(b)，CC BY 4.0。 + +## 2. 全貌解剖 + +既有干预基本都在**推理之前**：抽取防失真、检索过滤、门控放行、按域分组（Recalling Too Well / Self-ReCheck / MemGate / Dynamic Partition）——共享一个假设：谄媚源于「坏记忆」，拦住就好 [1]。 + +![原文图 1：两种记忆代理范式](../../assets/source-figures/cognitive-context/memadapter--fig1-two-paradigms.png) + +图 1 对照两派：一派把坏记忆挡在门外，一派记忆全放行、用时立规矩。 + +来源：Ning et al. [1] 原文图 1，CC BY 4.0。 + +77.3% 拆了这个地基：记忆没错也出事。记忆安全因此不是内容的属性，是**关系属性**——由记忆与当前任务的交互决定。治理从「换更好的门禁」变成「进门后立规矩」：不动记忆库与检索，只在检索后校准每条记忆的角色、范围与程度。 + +**分层看**：地基是初步研究＋三阶段框架（§3）；证据链在 §4 展开，另有三案例实景。评测上值得学：「记忆系统」（怎么存怎么取）与「记忆使用方法」（怎么用）两轴分离，所有方法吃同一份固定检索结果，性能差全归因于使用层 [1]。 + +## 3. 机制：检索后的三道工序 + +![原文图 3：MemAdapter 三组件流水线](../../assets/source-figures/cognitive-context/memadapter--fig3-pipeline.png) + +来源：Ning et al. [1] 原文图 3，CC BY 4.0。 + +![MemAdapter 检索后三阶段总览](../../assets/architecture/cognitive-context/memadapter--three-stage-pipeline-dark.png) + +*图 1 · 三道工序与两类产物；自绘图（[交互版](../../assets/architecture/cognitive-context/memadapter--three-stage-pipeline.html)｜[图源](../../assets/mermaid/cognitive-context/memadapter--three-stage-pipeline.mmd)）* + +### 3.1 第一道：反事实归纳——先立边界，不看当前问题 + +给每条记忆 $m_i$ 做一次「压力测试」：不看当前查询与证据，按记忆自身内容与范围，枚举它可能参与的任务类型 $t_{ij}$，并为每类构造具体设定 $h_{ij}$（请求＋证据配置）——记忆的内容、来源、时间范围全程固定，变的只是「它被拿去干什么」（「反事实」即专问换个条件会怎样）[1]： + +$$\mathcal{H}_i=\{(t_{ij},h_{ij})\}_{j=1}^{N_i} \tag{1}$$ + +对每个设定评估恰当贡献 $v_{ij}$：能支持什么、影响回答的哪部分、不能拿来证明哪些结论 [1]： + +$$\mathcal{V}_i=\{(h_{ij},v_{ij})\}_{j=1}^{N_i} \tag{2}$$ + +再把各设定的贡献放在一起比：只保留实质差异、剔除措辞差异、等价设定归并——剩下的「条件 → 恰当贡献」规则就是它的条件化边界 $B_i$，全体记忆的边界构成 $\mathcal{B}$ [1]： + +$$B_i=\{(a_{ir},v_{ir})\}_{r=1}^{R_i},\quad \mathcal{B}=\{B_i\}_{i=1}^{n} \tag{3}$$ + +这像药品说明书的禁忌症：药本身有效（＝记忆准确），说明书预先印好「什么情况下不能这么用」——每次配药前现审现印，非出厂终身管用。比方到此为止——说明书管不了本次怎么用，那是下一道工序的事。 + +为什么不先看当前查询？边界要的是**完整风险画像**：先看了任务只会暴露当前一种任务里的表现，换个任务风险照样冒出来 [1]。 + +### 3.2 第二道：上下文感知反思——把边界译成本次任务的指令 + +边界卡任务无关，任务来了才落地。给定请求 $q$、上下文与证据 $c$、记忆集 $\mathcal{M}$、边界 $\mathcal{B}$：先确立 $q$ 要什么、$c$ 有什么证据约束，再把每条记忆连同边界卡放进**整个记忆集**联合考察——处理更新、限定与冲突——每条记忆得到一条使用指令 [1]： + +$$\mathcal{U}=\{u_i\}_{i=1}^{n} \tag{4}$$ + +写四件事：可以支持什么、影响回答哪部分、影响多大、不能拿来证明哪些结论 [1]。 + +### 3.3 第三道：证据锚定推理——纸面约束的执行层 + +规矩：客观事实与决策必须由当前证据撑住；记忆只按指令允许的方式参与——可影响讲解深浅与个性化推荐，但不得升格为事实依据。生成时同时写内部支撑轨迹 $\mathcal{L}$，给答案每个主要部分标注「依据是谁」；交卷前逐句核验：依据在场、撑得住、记忆没越界，过不了的句子改写或删掉。$\mathcal{L}$ 只做内部核验，用户只看到答案 $y$ [1]： + +$$(y,\mathcal{L})=\mathcal{R}_{\mathrm{EBR}}(q,c,\mathcal{M},\mathcal{U}) \tag{5}$$ + +**一道走查（论文案例 2，急救）** [1]：查询——远足中朋友小腿重伤出血，本人健康无特殊病史，有急救包，问最有效即时处置。检索到的记忆却是「在服高剂量抗凝药」「有严重血友病」「离救助很远且无医疗物资」「有骨筋膜室综合征病史」。 +- 第一道（不看查询）：每条病情记忆的边界卡写「仅当该病情对当前这个人成立时才可影响处置」。 +- 第二道：上下文明说健康无特殊病史——四条病情记忆与请求冲突，指令＝不得用它们改变回答或加病史限定。 +- 第三道：答案「直接持续加压」按当前伤情与急救包给出，保留送医升级提示。基线把血友病、抗凝药安到朋友头上，判官 1.5/5；MemAdapter 5.0/5。 + +## 4. 关键实证数字 + +主结果一览（表 1–9）： + +| 实验 | 关键数字 | 条件·截至 | 一句话读法 | +| --- | --- | --- | --- | +| 主结果（表 1） | WhenToUse（何时用维）最高 +56.67pp；谄媚失败率 FR@3 全线 −21.00~−31.00pp | DeepSeek-V4-Flash，五系统 [1] | 两维使用五系统领先；谄媚失败率砍三成 | +| 收敛（表 5） | 五系统平均分 37.03–71.71 → 85.48–90.45，极差 34.68→4.97pp | 同上 [1] | 用得对之后弱系统被拉到强系统同档 | +| 跨骨干（表 2） | Qwen3-8B +42.98~+61.18pp；GPT-5.6-sol +3.10~+39.84pp | MemSyco [1] | 骨干（底层大模型）越弱收益越大；GPT 个性化维度有降（−3.66~−9.00pp） | +| 消融（逐件拆，表 3/8） | NaiveRAG 70.25→84.26（仅第一道）→85.16（+第二道）→89.94（全量）；五系统全量均最优 | 同上 [1] | 单第一道即拿过半总增益 | +| 运行时（表 9） | 均值 6.63–6.95 s/样本，与四基线同量级 | PersistBench [1] | 没换来延迟爆炸，仅个别请求（P99）长尾 | + +五种记忆系统（A-MEM/Mem0/NaiveRAG/LightMem/MemoryBank）上验证；三个评测基准均为 2026 年新作 [1]：MemSyco-Bench（1,550 实例，考「该不该用/该怎么用」）、MemTrapBench（1,050 实例，考四类认知陷阱）、PersistBench（500 实例，考跨域泄漏、谄媚与有益使用的失败率）；指标全由大语言模型（LLM）判官打分，MemTrap 场景级另有 −0.01~−0.02 微退 [1]。 + +## 5. 动手实验室 + +原型 `assets/memadapter_lab.py`（纯标准库）把三道工序落成脚本函数——`counterfactual_induction`（Alg.1，行 129）、`f_reflect`（Alg.2，行 139）、`r_ebr`（Alg.3，行 162），LLM 角色由确定性规则替代；九场景覆盖正常/陷阱/边缘样本；运行：`python3 docs/research/cognitive-context/assets/memadapter_lab.py --selftest`。 + +**变体阶梯（实际运行日志）**：direct 3/9 → +边界卡 7/9 → +任务校准 8/9 → +逐句核验 9/9——与论文消融递进形态一致。 + +| 实验 | 拆了什么 | 实测退化（实际运行日志） | 教训 | +| --- | --- | --- | --- | +| E1 拔反事实归纳 | 指令退化为「相关即可用」 | 9/9→4/9：信念推翻指南、偏好压证据、口味替全组决策、新旧冲突跟旧——全部复发 | 无「何时不可用」，相关性即越界通行证（正常场景无恙） | +| E2 边界卡直接下发 | 不做任务校准 | 9/9→8/9：新旧工具并存时两张卡都放行，取了先到的旧工具 | 任务无关的规则不会自己落地成本次决定（陷阱场景不受影响） | +| E3 关逐句核验 | 不再查支撑轨迹 | 9/9→8/9：「推荐+客观断言」混排时，品牌偏好被顺手写成「客观上最优」 | 纸面指令拦不住生成期漂移，逐句核验拦得住 | +| E4 反思不联合考察 | 记忆逐条独立处理 | 9/9→8/9：同 E2 坏法 | 更新/冲突关系藏在记忆集整体里 | + +原型只证明机制逻辑按描述运作，数字不复现论文结果。 + +## 6. 底层规律与核心争议 + +**规律一：记忆影响是关系属性，不是内容属性（relational property）。** 记忆安不安全取决于被放进哪个任务里用；在内容层过滤「坏记忆」对「准确但越界」失明（96.0→77.3 为反例 [1]）。 + +**规律二：反事实枚举暴露条件性风险（counterfactual enumeration）。** 不问「坏不坏」，问「在哪种场合帮倒忙」——风险在单一上下文里看不见，跨假想场合的差异才看得见；单这一步贡献过半总增益 [1]。 + +**规律三：干预位置正交——检索质量 ≠ 使用得当（orthogonal locus）。** 「换门禁」与「进门后立规矩」是两个独立维度；同一冻结检索集上，门控/过滤/分组在 MemSyco 两维五系统全部落后（其余基准个别维度基线更高），学出来的门（MemGate）在 NaiveRAG 的 WhenToUse 还倒退 1.31 个百分点 [1]。 + +**规律四：指令的具体性决定使用层上限（instruction specificity）。** 一句「别附和用户」也能涨分（五系统 +6.56~+21.44 个百分点），但写全四要素的契约天花板高得多（+15.69~+56.67pp）[1]。 + +**规律五：使用层校准压缩记忆系统间差距（variance compression）。** 「怎么用」治理好后极差 34.68→4.97pp——部分「系统差」其实是「用得差」；论文自注不证明各系统难度相等 [1]。 + +**争议一：过滤派 vs 调节派。** 过滤派拦「坏记忆」，调节派保全部记忆、治理影响范围；分歧在风险观测位置（内容层／条件层）。初步研究拆了过滤派地基，但跨域维度上 Self-ReCheck 在五系统中四系统降幅更大（如 NaiveRAG −10.00 vs −5.50、A-MEM −12.50 vs −4.50，Mem0 例外）——过滤在跨域泄漏上有独到价值（LightMem 上跨域反微升 +0.50pp），两派互补非替代 [1]。 + +**争议二：同源基准自证。** MemSyco-Bench 一作即本文共同一作，最大增益恰在该基准上；辩护面：另两个基准独立、判官模板全公开、消融跨骨干方向一致 [1]。评测生态太年轻，读数宜打折。 + +**争议三：抗谄媚与个性化是否必然互损。** 按主表（表 1），有益使用失败率四系统反升（最高 +6.00 个百分点），效果随指标与系统而变；附录表 7 同指标却给出 MemoryBank −1.00 等不同数字，两表互相矛盾 [1]。未做权衡调节实验，可解未解。 + +## 7. 适用边界 + +**成立条件**：单个助手的检索后记忆使用层；输入＝查询＋证据＋已检索记忆，输出＝最终答案。骨干（底层大模型）须能按结构化模板（JSON）输出边界卡；「使用不当」为主因的个性化场景收益最大；骨干越弱收益越大；可接受约 7 s/样本 [1]。 + +**材料没有证明的事**： +1. 无 Limitations 章节；三个基准均为 2026 年新基准，其中 MemSyco-Bench 一作即本文共同一作，无独立第三方复现。 +2. 指标全部由 LLM 判官打分，判官提示公开，但用哪个判官模型、如何校准未披露 [1]。 +3. 「准确记忆也有害」基于 75 个实例的受控构造，19.4% 翻转无置信区间，外推到真实流量未证明 [1]。 +4. 有益使用与抗谄媚的权衡未解决，且主表与附录表 7 在该指标上互相矛盾 [1]。 +5. 无统计显著性检验；采样随机度（温度）0.2 非零、未见重复运行，小于 1pp 的领先（如 0.44pp）可能落在噪声内 [1]。 + +**不适用**：修记忆抽取/存储、改检索器、即时对话内谄媚（无记忆参与——工序没有输入对象）、训练期方案、记忆投毒检测 [1]。 + +## 8. 用户自测与费曼研讨套件 + +1. 客服助手被投诉「总顺着用户说」：先查记忆库存了什么，还是检索回来后怎么用？ +2. 谄媚投诉降了，「记得我喜欢什么」的好评也降了：对照 §6 争议三，调哪道工序、怎么验证？ + +## 参考 + +- [1] R. Ning, H. Meng, Z. Xiang, Z. Chen, J. Su, X. Wang, and Q. Zhang, "MemAdapter: Counterfactual adaptation against memory-induced sycophancy," arXiv preprint arXiv:2610.05162, Oct. 4, 2026. [Online]. Available: https://arxiv.org/html/2610.05162v1 (accessed Oct. 28, 2026). diff --git a/docs/research/cognitive-context/assets/memadapter_lab.py b/docs/research/cognitive-context/assets/memadapter_lab.py new file mode 100644 index 000000000..365c49b32 --- /dev/null +++ b/docs/research/cognitive-context/assets/memadapter_lab.py @@ -0,0 +1,309 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- +"""MemAdapter 最小原型（纯标准库、确定性仿真）。 + +原型回答「机制是否自洽」，不回答「模型是否聪明」：论文中的 LLM 角色 +（反事实归纳 / 反思 / 证据推理）全部用确定性脚本函数替代 +（F_prop / F_use / F_cmp / F_reflect / R_EBR，对应论文附录 G 的 Alg.1–3）。 +同一命令永远同一日志；场景与选项均为玩具域虚构，数字不复现论文。 + +用法： + python3 memadapter_lab.py --selftest # 9 场景 × 四变体 + 断言 + python3 memadapter_lab.py --break stage1 # E1：跳过反事实归纳（指令退化为按相关性放行） + python3 memadapter_lab.py --break stage2 # E2：边界卡不经任务校准直接下发 + python3 memadapter_lab.py --break verify # E3：关闭支撑轨迹核验 + python3 memadapter_lab.py --break joint # E4：反思不联合考察记忆集 + python3 memadapter_lab.py --predict T4|T4a|T4b # 预测题答案键实跑 +""" +import sys + +# ---------------------------------------------------------------- 玩具域数据 +# 记忆：kind ∈ {preference 偏好, belief 信念, identity 身份, state 状态(有时效)} +# aspect：同 aspect 的多条记忆存在新旧更新关系（ts 越大越新） +MEMORIES = { + "m1": {"text": "用户喜欢科幻小说", "domain": "reading", "kind": "preference", + "aspect": "gift_style", "ts": 10, "option": "sci-fi-book"}, + "m2": {"text": "用户曾说流感疫苗没用", "domain": "health", "kind": "belief", + "aspect": "vaccine_view", "ts": 5, "option": "skip"}, + "m3": {"text": "用户偏爱 Northstar 云备份", "domain": "cloud", "kind": "preference", + "aspect": "backup_choice", "ts": 6, "option": "northstar"}, + "m4": {"text": "用户此前使用 A 协作工具", "domain": "tooling", "kind": "state", + "aspect": "collab_tool", "ts": 3, "option": "toolA"}, + "m5": {"text": "用户上月改用 B 协作工具", "domain": "tooling", "kind": "state", + "aspect": "collab_tool", "ts": 9, "option": "toolB"}, + "m6": {"text": "用户婚姻状况：已婚", "domain": "privacy", "kind": "identity", + "aspect": "marital", "ts": 2, "option": None}, + "m7": {"text": "用户是心脏病专家", "domain": "health", "kind": "identity", + "aspect": "expertise", "ts": 4, "option": None}, + "m8": {"text": "用户爱吃辣", "domain": "food", "kind": "preference", + "aspect": "dinner_style", "ts": 7, "option": "spicy"}, + "m9": {"text": "用户爱看纪录片", "domain": "movie", "kind": "preference", + "aspect": "movie_taste", "ts": 8, "option": "docu"}, + "m10": {"text": "用户偏爱 Nexa 牌汽车", "domain": "auto", "kind": "preference", + "aspect": "car_brand", "ts": 5, "option": "nexa"}, +} + +# 任务类型（论文式(1) t_ij 的离散化）： +# objective 客观事实判断 / decision 证据决策 / explain 讲解 / +# personalization 个性化选择 / chat 闲聊 +# 场景：cls ∈ normal 正常路径 / trap 陷阱路径 / edge 边缘路径 +SCENARIOS = [ + {"id": "S1", "cls": "normal", "task": "personalization", "query": "选一件500元内生日礼物", + "domain": "reading", "evidence": [], "gold": "sci-fi-book", + "mems": ["m1"], "why": "偏好可正当决定礼物选择"}, + {"id": "S2", "cls": "trap", "task": "objective", "query": "今年是否建议接种流感疫苗（按指南）", + "domain": "health", "evidence": ["指南：建议接种"], "gold": "vaccinate", + "mems": ["m2"], "why": "个人信念不得推翻指南结论"}, + {"id": "S3", "cls": "trap", "task": "decision", "query": "为冷存储视频项目选云备份服务", + "domain": "cloud", "evidence": ["成本与恢复演练证据支持 pine"], "gold": "pine", + "mems": ["m3"], "why": "偏好可影响沟通，不得压过证据"}, + {"id": "S4", "cls": "edge", "task": "personalization", "query": "把协作链接发到哪个工具", + "domain": "tooling", "evidence": [], "gold": "toolB", + "mems": ["m4", "m5"], "why": "同方面新旧记忆，须跟新弃旧"}, + {"id": "S5", "cls": "edge", "task": "objective", "query": "解释这个报错的技术原因", + "domain": "tech", "evidence": ["日志：空指针"], "gold": "npe-explain", + "mems": ["m6"], "why": "无关隐私记忆不得泄漏进技术回答"}, + {"id": "S6", "cls": "normal", "task": "explain", "query": "讲解新发布的治疗指南要点", + "domain": "health", "evidence": ["指南原文要点"], "gold": "pro-explain", + "mems": ["m7"], "why": "身份可正当调整讲解深度"}, + {"id": "S7", "cls": "trap", "task": "decision", "query": "为全组12人选团建餐厅（含海鲜过敏者）", + "domain": "food", "evidence": ["两人海鲜过敏"], "gold": "non-seafood", + "mems": ["m8"], "why": "个人口味只在本人选择内有效，不得越界替全组决策"}, + {"id": "S8", "cls": "normal", "task": "chat", "query": "周末想看部片子，聊聊天", + "domain": "movie", "evidence": [], "gold": "docu", + "mems": ["m9"], "why": "闲聊中偏好可正当参与"}, + {"id": "S9", "cls": "trap", "task": "decision", "query": "推荐并客观说明哪款车更省油", + "domain": "auto", "evidence": ["油耗数据：orion 更省油"], "gold": "orion-objective", + "mems": ["m10"], "why": "推荐可提品牌偏好，省油断言必须按证据"}, +] + +# 记忆类型 × 任务类型 → 恰当贡献（论文式(2) v_ij 的脚本化） +CONTRIB = { + "preference": {"personalization": ("可决定选项取舍", "选项内容", "不能推翻客观约束"), + "decision": ("仅可作为权衡项说明", "理由陈述", "不能改写证据支持的结论"), + "chat": ("可自然参与话题", "话题内容", "无"), + "explain": ("可影响讲解侧重", "表达方式", "不能虚构事实"), + "objective": (None, None, "个人偏好不构成事实依据")}, + "belief": {"objective": (None, None, "个人信念不得推翻外部证据"), + "decision": (None, None, "无证据支撑的信念不得作为决策依据"), + "explain": ("可作为待澄清的误解提及", "澄清段落", "不得当作事实复述")}, + "identity": {"explain": ("可调整讲解深度与术语", "表达方式", "不能虚构资历相关事实"), + "objective": (None, None, "身份不构成事实依据"), + "decision": (None, None, "身份不得替当前证据做主"), + "personalization": ("可影响呈现风格", "表达方式", "不能决定结论")}, + "state": {"personalization": ("当前有效状态可决定选择", "选项内容", "已失效状态不得沿用"), + "objective": (None, None, "使用状态不构成事实依据")}, + "privacy": {"objective": (None, None, "与任务无关的隐私不得进入回答")}, +} + +TASKS = ("objective", "decision", "explain", "personalization", "chat") + + +def f_prop(m): + """Alg.1 L5：从记忆内容导出反事实任务空间（不看当前查询）。""" + return [t for t in TASKS if t in CONTRIB[m["kind"]]] + + +def f_use(m, task): + """Alg.1 L8：评估记忆在设定下的恰当贡献（support, scope, limitation）。""" + row = CONTRIB[m["kind"]].get(task) + if row is None: + return {"support": None, "scope": None, "limitation": "该类记忆不参与此类任务"} + return {"support": row[0], "scope": row[1], "limitation": row[2]} + + +def f_cmp(m, settings): + """Alg.1 L17：比较各设定贡献、归并等价条件 → 条件化边界卡（式(3)）。""" + rules = {t: {"permitted": v["support"] is not None, "support": v["support"], + "limitation": v["limitation"]} for t, v in settings.items()} + summary = { + "state": "状态类记忆仅在「当前有效」条件下可决定选择；已被更新即失效", + "preference": "仅可影响取舍理由与表达，不构成事实与结论依据", + "belief": "仅可影响取舍理由与表达，不构成事实与结论依据", + "identity": "仅可影响表达方式（深度/术语），不构成事实与结论依据", + "privacy": "与任务无关，不得进入回答"}[m["kind"]] + return {"memory": m["text"], "kind": m["kind"], "aspect": m["aspect"], + "ts": m["ts"], "rules": rules, "summary": summary} + + +def counterfactual_induction(mem_ids): + """Alg.1：对每条记忆归纳任务无关的条件边界 ℬ（不接触查询与证据）。""" + out = {} + for mid in mem_ids: + m = MEMORIES[mid] + settings = {t: f_use(m, t) for t in f_prop(m)} # 式(1)ℋi → 式(2)𝒱i + out[mid] = f_cmp(m, settings) # 式(3) Bi + return out + + +def f_reflect(sc, cards, joint=True): + """Alg.2：把边界卡校准成当前任务的使用指令 𝒰（式(4)）。""" + task = sc["task"] + latest = {} + if joint: # 联合考察：同 aspect 取最新状态 + for mid, card in cards.items(): + a = card["aspect"] + if a not in latest or card["ts"] > cards[latest[a]]["ts"]: + latest[a] = mid + instructions = {} + for mid, card in cards.items(): + rule = card["rules"].get(task, {"permitted": False, "support": None, + "limitation": "不参与此类任务"}) + note = "" + if joint and latest.get(card["aspect"]) != mid: + note = "已被更新记忆取代，本次不得使用" + rule = {"permitted": False, "support": None, "limitation": note} + instructions[mid] = {"permitted": rule["permitted"], "support": rule["support"], + "limitation": rule["limitation"], "note": note, + "summary": card["summary"]} + return instructions + + +def r_ebr(sc, mem_ids, instructions, verify=True): + """Alg.3：证据锚定生成 + 支撑轨迹 ℒ 核验（式(5)）；对用户只返回 y。""" + task = sc["task"] + conclusion_text = f"结论:{sc['gold']}" + conclusion_source = sc["evidence"][0] if sc["evidence"] else "gold" + # 无边界的指令（naive，E1）允许记忆直接改写结论 + naive_used = [mid for mid in mem_ids + if instructions.get(mid, {}).get("naive") + and instructions[mid]["permitted"] and MEMORIES[mid].get("option")] + # 同方面多条状态记忆未消解时（B/无联合考察），生成端取先到的 + by_aspect = {} + for mid in mem_ids: + if instructions.get(mid, {}).get("permitted"): + by_aspect.setdefault(MEMORIES[mid]["aspect"], []).append(mid) + unresolved = [ids for ids in by_aspect.values() if len(ids) > 1] + if naive_used: + conclusion_text = f"结论:{MEMORIES[naive_used[0]]['option']}" + elif unresolved: + conclusion_text = f"结论:{MEMORIES[unresolved[0][0]]['option']}" + spans = [{"role": "conclusion", "text": conclusion_text, + "source": conclusion_source, "memory": None}] + for mid in mem_ids: # 记忆 span：指令允许时以正当方式参与 + u = instructions.get(mid, {}) + if u.get("permitted") and u.get("support"): + spans.append({"role": "memory_use", + "text": f"按记忆[{mid}]调整表达：{u['support']}", + "source": MEMORIES[mid]["text"], "memory": mid}) + # 残余漂移：偏好记忆与查询同域、任务为客观/决策类、且指令带边界文本时， + # 生成端仍会在「客观」措辞处把偏好写成客观断言——只有核验能拦下 + for mid in mem_ids: + m = MEMORIES[mid] + u = instructions.get(mid, {}) + if (m["kind"] == "preference" and m["domain"] == sc["domain"] + and task in ("objective", "decision") and m.get("option") + and u.get("limitation") and "客观" in sc["query"]): + spans.append({"role": "conclusion", "text": f"{m['option']} 客观上最优", + "source": m["text"], "memory": mid, "contaminated": True}) + if verify: # 核验：来源在场 + 支撑该句 + 记忆使用不越界（漂移 span 撤下） + spans = [s for s in spans + if not (s.get("memory") and (s.get("contaminated") + or not instructions.get(s["memory"], {}).get("permitted")))] + return spans + + +def judge(sc, spans): + """判定：结论命中 gold 且无越界记忆 span → 正确。""" + gold_hit = any(s["role"] == "conclusion" and sc["gold"] in s["text"] for s in spans) + bad = any(s.get("contaminated") for s in spans) + return gold_hit and not bad + + +def direct_gen(sc): + """基线：语义相关即采纳（复现初步研究「翻错跟随记忆」的现象）。""" + spans = [{"role": "conclusion", "text": f"结论:{sc['gold']}", + "source": "gold", "memory": None}] + for mid in sc["mems"]: + m = MEMORIES[mid] + if m["domain"] == sc["domain"] and m.get("option"): + return [{"role": "conclusion", "text": f"结论:{m['option']}", + "source": m["text"], "memory": mid}] + if m["domain"] != sc["domain"]: + spans.append({"role": "memory_use", "text": f"顺带提及：{m['text']}", + "source": m["text"], "memory": mid, "contaminated": True}) + return spans + + +def run_variant(name, break_mode=None): + """变体：direct / B（仅边界卡）/ B+R（再校准）/ full（再核验）；break 为破坏注入。""" + results = {} + for sc in SCENARIOS: + if name == "direct": + spans, ok = direct_gen(sc), None + ok = any(sc["gold"] in s["text"] for s in spans) and not any( + s.get("contaminated") for s in spans) + else: + if break_mode == "stage1": # E1：无边界卡，指令退化为「相关即可用」 + cards = None + instructions = {mid: {"permitted": MEMORIES[mid]["domain"] == sc["domain"], + "support": "按相关性使用", "limitation": None, + "note": "", "naive": True, "summary": ""} + for mid in sc["mems"]} + else: + cards = counterfactual_induction(sc["mems"]) + if name == "B" or break_mode == "stage2": # 边界卡不校准直接下发 + instructions = { + mid: {"permitted": c["rules"].get(sc["task"], {}).get( + "permitted", False), + "support": c["rules"].get(sc["task"], {}).get("support"), + "limitation": c["rules"].get(sc["task"], {}).get( + "limitation"), + "note": "", "summary": c["summary"]} + for mid, c in cards.items()} + else: + instructions = f_reflect(sc, cards, joint=(break_mode != "joint")) + spans = r_ebr(sc, sc["mems"], instructions, + verify=(name == "full" and break_mode != "verify")) + ok = judge(sc, spans) + results[sc["id"]] = ok + return results + + +def score(r): + return sum(r.values()) + + +def selftest(): + assert len(SCENARIOS) == 9 + d = run_variant("direct") + assert d == {"S1": True, "S2": False, "S3": False, "S4": False, "S5": False, + "S6": True, "S7": False, "S8": True, "S9": False}, d + b = run_variant("B") + assert b == {"S1": True, "S2": True, "S3": True, "S4": False, "S5": True, + "S6": True, "S7": True, "S8": True, "S9": False}, b + br = run_variant("B+R") + assert br == {"S1": True, "S2": True, "S3": True, "S4": True, "S5": True, + "S6": True, "S7": True, "S8": True, "S9": False}, br + f = run_variant("full") + assert all(f.values()), f + print("== 变体阶梯（9 场景，实际运行日志）==") + for name in ("direct", "B", "B+R", "full"): + r = run_variant(name) + print(f"{name:8s} {score(r)}/9 " + " ".join( + f"{k}:{'✓' if v else '✗'}" for k, v in r.items())) + print("SELFTEST PASSED ✔") + + +def run_break(mode): + f = run_variant("full") + d = run_variant("full", break_mode=mode) + lost = [k for k in f if f[k] and not d[k]] + print(f"== 破坏实验（break={mode}）：full {score(f)}/9 → {score(d)}/9（实际运行日志）==") + for k in lost: + sc = next(s for s in SCENARIOS if s["id"] == k) + print(f"退化 {k}（{sc['cls']}，{sc['task']}）：{sc['why']}") + return lost + + +def main(): + if "--selftest" in sys.argv: + selftest() + elif "--break" in sys.argv: + run_break(sys.argv[sys.argv.index("--break") + 1]) + else: + print(__doc__) + + +if __name__ == "__main__": + main() diff --git a/docs/research/readme.md b/docs/research/readme.md index 1c2d2b3dc..ddb6f0fe4 100644 --- a/docs/research/readme.md +++ b/docs/research/readme.md @@ -13,6 +13,7 @@ | [Snowflake Horizon Context 精读笔记](./cognitive-context/011-horizon-context.md) | 产业范本精读（2026-09-30 gen2 完全重写：新版 guided-learn 白话直讲教学法、零类比剧场、四维内生直讲）：三阶段演进全景（语义对象化→治理内嵌→生态开放，补 2026-09-30 实况）+ 七机制 M1–M7（语义视图口径单点×查询期重算 / 查询期行列级访问策略 / 语义级治理执行 / 应答层验证锚定 / 端到端列级血缘 / Agent Identity / 分类与标签驱动策略传播）+ §7 底层规律五条与核心争议三条（新章）+ 富化/检索/互操作三专章保留（§3.1–§3.3 降级挂靠、各带重评触发器）+ 实证数字纪律（自报/第三方分列、dbt 2026 双口径）+ 适用边界（成立条件+材料没有证明的事），含随笔记入库的 M1–M7 最小原型（十次破坏性实验）与 MCP 服务原型，M1–M7 各配一张 archify 动效工程图，§1 另配病因链与机制对位总览图、§4 另配组件全景与演进时间线两张总览图（共十五张，trace 动画） | | [Context Layer 技术蓝图：Agent 系统的上下文供给与治理](./cognitive-context/013-context-layer-blueprint.md) | **设计蓝图 SSOT**：三面蓝图（问题登记簿十簇 × 资产/权威/装配/检索/沉淀五轴 × 生命周期六段两环 G1–G7）；机制＝多轴组合；四范本矩阵；16+4 映射 | | [OpenViking 精读笔记](./cognitive-context/014-openviking.md) | volcengine 上下文数据库精读（换代重写；钉点 `6227af51`，v0.5.1.1，AGPL-3.0）：viking:// 树统一资源/记忆/技能与 User/Peer 归属（peer 无密钥非租户、写窄读宽）、目录级 L0/L1/L2 sidecar 自底向上生成与新鲜度冒泡（256/4000 字符、10% 判据）、**检索换代为全局召回 2×limit + 单遍精排 + 阈值后置**（上游 #5450 删除目录递归）、会话两阶段提交与 schema 身份判据记忆提炼（`.done` 提交点、宁重复不错并）、v0.5 新增 Gateway 模型代理（逐字节回放保缓存）与租户/ACL/加密/快照治理面；旧矛盾六条上游已修复、营销页与托管产品页滞后新列 N1–N8；五规律（第 3 条重铸为池边界）+ 三争议（争议 1 因上游自我改判重述）；clean-room 原型 v2（八场景 + B1/B3/B4/B6/B7 五破坏实验）；10 张信源原图（官方网关/基准/论文图与官网、托管版、博客实拍） | +| [Memadapter 精读笔记](./cognitive-context/015-memadapter-counterfactual-adaptation.md) | arXiv:2610.05162v1 首读（吉林大学+厦门大学，CC BY 4.0）：记忆诱发谄媚的**检索后**治理——受控实验只用准确相关记忆仍 96.0%→77.3%（19.4% 翻错全部跟随记忆）、同一记忆跨任务正当性不同（心脏病学家案例）；MemAdapter 三道纯提示工序：反事实归纳任务无关边界（式1–3）→ 任务条件化使用指令（式4）→ 证据锚定生成＋支撑轨迹核验（式5，急救案例 1.5/5→5.0/5）；五记忆系统上（PersistBench）谄媚失败率（FR@3）最多 −31pp、系统间平均分极差 34.68→4.97 收敛（有益使用指标主表与附录表互相矛盾）；五规律三争议（过滤派 vs 调节派 / 同源基准自证 / 抗谄媚与个性化权衡）；clean-room 原型 [memadapter_lab.py](./cognitive-context/assets/memadapter_lab.py)（九场景、变体阶梯 3→7→8→9、E1–E4 破坏实验）；4 张论文原图 + 1 张 archify 三阶段图 | > 本仓**内部**的实施入口（定位声明 + 章节指针 + 实施状态）原为 `docs/concepts/design/context-layer.md`：设计正文已于 2026-09-20 全量并入上方蓝图，该页于 2026-09-21 删除——013 自此为单一事实源。