deepmemory — Long-term Memory for DeepSeek Harness

中文版

Give your DeepSeek Harness agents cross-session memory and a near-infinite context. Facts, preferences, decisions, plans and tasks survive across sessions — structured, searchable and protected. Implemented natively for DSH: a Python memory backend, a Cordis web plugin and three agent presets.

Highlights

• 5-class memory model — semantic, short-term, process, source archive, compressed archive
• 3 storage tiers — active → cold → archive, with automatic demotion and lifecycle decay
• Sensitivity-aware by default — PII and natural-language password detection, redaction at write time, approval-gated reveal, full audit trail
• Decision lifecycle — proposed → exploring → pending → adopted / rejected / superseded / invalid, so conflict resolution is explicit instead of silent overwrite
• Hybrid retrieval — BM25 + vector + graph, RRF fusion, recency × importance × relevance weighting
• Dual domains — work / life memory separated at write and query time
• Mode presets — two production presets (task / daily) plus an extension template
• WebUI — memory panel, entity graph, archive, maintenance, per-session config
• Dynamic task board — 任务看板已独立为 dsh-livetaskboard 插件（状态机 + 外援 sol/子代理），本仓库保留任务系统接口
• Explicit ownership — tasks are Workspace-scoped and link to a Session; state cards are independently versioned per Session

Memory Model (brief)

raw conversation → memory entry (class + domain + actor + sensitivity)
                → atoms (TTL / decay / reinforcement)
                → graph (entities + relations)

• Classes: semantic (stable facts), short_term (recent daily context), process (task-progress details), source_archive (original excerpts), compressed_archive (consolidated summaries).
• Tiers: memories start in active, demote to cold after their window (7 days short-term / 15 days process by default), and live in cold for a year before archive. Recall prefers active, then cold.
• Sensitivity: normal / sensitive / protected / secret. Detectors cover bank cards (Luhn), Chinese ID numbers (checksum), phones, API keys/tokens, and Chinese natural-language passwords. Matches are redacted on write; revealing original text requires approval (3 attempts, 30-min TTL, audited).
• Decisions: a status machine prevents a later conversation from silently overriding an earlier decision. Rejected alternatives are downgraded; only the user can mark a plan invalid.
• Retrieval: three-way RRF retrieval with a cache, per-domain filtering, actor filtering and sensitivity filtering.

Mode Presets — pick your agent's shape

deepmemory ships as three presets, each a complete agent.cordis.yml configuration. Choose one per session.

Preset	Shape	Includes	Skips
task 任务工作模式	full coding agent	all tools, plan mode, sub-agents, workflow, process memory, budget profile task-default; 任务看板由 dsh-livetaskboard 提供	—
daily 日常问答模式	lightweight Q&A	web search, short-term memory, daily state card, topic continuity, budget profile daily-default	sub-agent orchestration, workflow
blank-template	extension template	minimal persona, optional-tool comments, plugin/ entry point, preset-local realm example, budget comments	deepmemory injection, extraction, state card

Each preset declares a budget contract: budget_profile + priority_allocation (per-component priority and min_tokens). Unused budget returns to the pool and is redistributed by priority — the model context window is the hard cap, components negotiate within it.

Quick Start

git clone https://github.com/UnKnownFish125/dsh-deepmemory.git && cd dsh-deepmemory
sudo bash scripts/install.sh

Idempotent and safe to re-run. It installs the memory backend (memory-server, systemd unit, health check), the web plugin (dsh-deepmemory bundle registered in profiles/web/package.json, client.js auto-converted to __ModuleLoader__ format), and the agent presets under ${DSH_HOME}/.agent-presets/. The script never restarts the DSH web process itself — it prints a restart checklist and leaves that to the administrator.

Override environment: DSH_HOME, APP_DIR, VENV_PY.

Embedding — local model or API (pick one)

Semantic retrieval runs on a pluggable embedding provider, configured under the embedding group (rendered automatically in the WebUI「配置」tab, or via the session config API):

Key	Default	Meaning
embedding.provider	local	local = fastembed inference on this machine; api = any OpenAI-compatible /v1/embeddings endpoint
embedding.local_model	BAAI/bge-small-zh-v1.5	Local model; downloaded automatically from the HF mirror on first use
embedding.api_base_url	—	Required when provider=api, e.g. https://api.openai.com/v1
embedding.api_key	—	Required when provider=api; or inject via EMBED_API_KEY env var to keep it out of the DB
embedding.api_model	text-embedding-3-small	Model name for the API provider

Default model: BAAI/bge-small-zh-v1.5 — 512-dim, tuned for Chinese, ~30 MB ONNX, fully offline after first download (served via the HF mirror hf-mirror.com; override with HF_ENDPOINT).

Switching providers (e.g. local → api) is safe: the FAISS index is rebuilt automatically with the new dimension, existing memories are re-embedded on demand, and no data is lost.

Usage

1. Choose a preset — start a new session and select 任务工作模式 (task) or 日常问答模式 (daily). Memory plugin loads automatically; no dynamic define+run needed.
2. WebUI — open the「记忆」tab in the conversation view: manage memories, inspect the entity graph, edit the current conversation card, and use the current Workspace task kanban. Every task can open or rebind its linked conversation.
3. API — HTTP backend on :6230: v2 task API at /v1/v2/tasks?workspace_id=...（任务看板 UI 由 dsh-livetaskboard 消费）, session cards at /v1/v2/cards/<kind>/<session_id>, recall/lifecycle under /v1/v2/, session config at /v1/config/session/*, and sensitivity audit at /v1/sensitive/audit.

Extending

• Child plugin (recommended) — copy agent-preset/blank-template/ to ${DSH_HOME}/.agent-presets/<your-preset>/, rename it, put your business code in plugin/. The template deliberately ships without deepmemory so you start from a clean slate; wire in the pieces you need (memory injection, state card, budget profile).
• Custom preset — base it on task/ or daily/, adjust persona, tool catalog and budget_profile; keep session state out of preset files and in plugin code.
• Memory classes & API — memory-server/v2_domain.py is a stdlib-only data-contract layer; extend constants (MEMORY_CLASSES, STORAGE_TIERS, …) and lifecycle primitives there, then expose routes in server.py.
• Sensitivity rules — add or tune detectors in memory-server/sensitive.py; unit tests live in memory-server/tests/test_sensitive.py.

Architecture

┌─ write ────────────────────────────────┐
│ ① cheap LLM extraction (turn-stopping) │
│ ② memory_save model tool (explicit)    │
│ ③ WebUI manual entry                   │
└────────────────┬───────────────────────┘
                 ▼
   memory-server (Python, systemd :6230)
   SQLite + FAISS + BM25 (jieba)
   RRF fusion + tri-factor weighting
   sensitive redaction / decision states / lifecycle decay
                 ▼
┌─ recall injection ─────────────────────┐
│ silent system-message injection        │
│ [state card] + [Top-K memories]        │
│ scopes: session / workspace / global   │
└────────────────────────────────────────┘

Inspiration & References

• Generative Agents: Interactive Simulacra of Human Behavior — memory stream, recency × importance × relevance retrieval, reflection
• Memory in the Age of AI Agents: A Survey — 2025 survey of agent memory architectures
• HiMem: Hierarchical Long-Term Memory for LLM Long-Horizon Agents — hierarchical long-term memory

License

AGPL-3.0. This project is an independent native implementation for DeepSeek Harness, not a code port. Modifications, derivatives and distribution (including network service provision) must comply with AGPL terms.