Skip to content

experience-memory.md

Latest commit

 

History

History
75 lines (53 loc) · 3.84 KB

experience-memory.md

File metadata and controls

75 lines (53 loc) · 3.84 KB
Raw

Experience Memory

GraphFlow’s Experience layer turns storage → reflection → experience: raw episodes and graph facts are reflected into scored skills and lessons, then retrieved as organizational memory under the same context budget as code.

Related: Context contract · Team memory security · Flywheel auto-capture

Framing: Storage → Reflection → Experience

Stage What lives here How it advances
Storage Graph nodes (File/Symbol/Module), Decision episodes, Skill payloads, optional artifacts Index, report_outcome, auto-capture hooks
Reflection Score updates, outcome taxonomy (proven / correctable / anti-pattern / noise), SkillOpt-lite guidance edits, consolidation proposals Flywheel after pass/fail; diagnose / skill report
Experience What agents actually reuse: skill hints on plans, episode recall, L3 anchors in context packages graphflow_context / plan / skill_insights under token budget

Without reflection, storage is only a log. Without retrieval under a budget, reflection never becomes experience.

Organizational memory building blocks

Artifact Role
Episodes Task → outcome → lessons (and optional deviation). Feed recall and skill learning.
Skills Distilled, scored patterns with optional guidance and provenance / canary gates.
Graph artifacts Portable gzip/JSON snapshot of the knowledge graph (graphflow artifact export/import).
Memory pack (Markdown) Human-readable skills + recent episodes for review and onboarding (artifact export-memory).
Skill sync package Machine mergeable team skills (.graphflow/skills/…) with security gates.

Together they form engineering memory that can move with the repo — reviewable like code, gated against poisoning (see team-memory-security.md).

Export / import

Binary / JSON graph artifact

graphflow artifact export                      # → graphflow-out/graphflow-graph.artifact.gz
graphflow artifact export --include-episodes   # keep episode Decision nodes
graphflow artifact import [path]

MCP: graphflow_artifact with mode: "export" | "import".

Markdown experience memory pack

Human-readable companion (not a full graph restore):

graphflow artifact export-memory
# → graphflow-out/memory-pack/README.md
# → graphflow-out/memory-pack/skills.md
# → graphflow-out/memory-pack/episodes.md

graphflow artifact export-memory path/to/dir
graphflow --json artifact export-memory

Programmatic: exportExperienceMemoryPack(config, outputDir?) in src/graph/memory-pack.ts.

Skill package (team MERGE)

graphflow skill sync export
graphflow skill sync import   # newer updatedAt wins; --force overwrites

Agent Plugin vs graphflow install

Path Primary use What you get
Agent Plugins 1.0 (plugin.json / mcp.json / skills/) Preferred single-host install (e.g. Cursor local plugins / marketplace) MCP + canonical Skill discovered together
DeepSeek Harness bundle (dsh.bundle + cordis.patch.yml + dsh/plugin.mjs) dsh plugin add @roarpeng/graphflow then npx @deepseek-ai/dsh web MCP (mcp__graphflow__*，cwd=工作区）+ runtime skill + agent/disposed 飞轮关闭。无 VS Code UI
npx @roarpeng/graphflow install Fallback / multi-agent / Rules MCP + Skill + Rules across 15+ detected agents (incl. ~/.dsh)
graphflow doctor Diagnostics Which agents are installed / wired

Rules and multi-agent wiring still use install when the host does not consume Agent Plugins. Context packaging itself is unchanged — always start with graphflow_context (context-contract.md).