@wwskills/dsh-long-memory

Long-term cross-session memory + self-evolving learning plugin for DeepSeek Harness.

SQLite-backed, FTS5 + optional embedding recall, append-only audit log, L7 auto-extraction, lesson capture from user corrections, rule lifecycle with context injection. Single-bundle dual-face packaging (Node service + browser UI).

Features

Memory & Recall

• 8 mem_* tools — search, record, status, stats, forget, confirm, scope list, scope set
• FTS5 full-text search with CJK support (works out of the box, zero config)
• Optional embedding — Ollama (local, zero cost) or any OpenAI-compatible API
• Hybrid recall — BM25 + vector + RRF fusion when embedding is enabled, trust-weighted ranking
• L7 auto-extraction — automatically extracts memories from conversations on turn/end using your DSH LLM provider (zero extra config)
• Keyword fallback — if LLM is unavailable, regex-based keyword extraction kicks in
• Confirm queue — low-confidence L7 extractions go to confirm queue for user approval
• Supersession — overlapping L7 extractions auto-mark old memories as superseded
• File tracks — MEMORY.md session markers + memory/YYYY-MM-DD.md daily notes
• Audit log — append-only, tracks every memory operation

Self-Evolving Learning (merged from dsh-agent-evolve)

• Signal word detection — real-time capture when user says "不对" / "wrong" / "should be" etc. (8 CN + 7 EN, configurable)
• Tool error capture — tools/result event listener auto-records tool failures as corrections
• Agent error capture — agent/error event listener auto-records harness-level errors
• Lesson extraction — LLM-powered structured lesson extraction from correction-triggering messages
• Rule lifecycle — proposed → approved → rejected → archived → promoted_to_agents
• Rule injection — approved rules auto-injected into agent context via agent/pre-step (≤800 token budget, hit_count tracking)
• Rule conflict detection — Jaccard overlap >60% warns on approve
• AGENTS.md promotion — high-hit-count rules can be promoted to AGENTS.md format
• Daily decay — stale rules (90 days unhit) auto-archived
• Persona building — auto-builds user persona from USER-type memories (tech stack, coding style, communication, common tasks)

Browser UI

• Sidebar + Main layout — resizable sidebar (160-360px) with scope filter + search + stats
• 4 Tab management — 教训 (Corrections) / 规则 (Rules) / 记忆 (Memories) / 画像 (Persona)
• Recall test panel — keyword-based memory recall testing with score/source display
• Settings Modal — extraction, persona, embedding, and signal word config in a modal dialog
• Toast notifications — success/error/warning feedback for all operations
• PopoverMenu — hover-revealed action menu (archive/delete) on memory cards
• 30s auto-refresh — stats and badges stay current

Install

dsh plugin --profile web add @wwskills/dsh-long-memory

Configuration

Defaults are sensible. Override via your profile's patch layer as needed.

Embedding

Provider	Use case	Cost
none	FTS5 keyword only (default)	Zero
ollama	Local Ollama service	Zero (local)
openai-compatible	Any OpenAI-style API	Per-call

- id: long-memory
  config:
    embedding:
      provider: 'ollama'          # 'none' | 'ollama' | 'openai-compatible'
      model: 'bge-m3'
      dimension: 1024
      ollama:
        base_url: 'http://127.0.0.1:11434'

L7 Auto-extraction

L7 reads your DSH LLM provider config automatically — no extra API key needed.

- id: long-memory
  config:
    l7:
      enabled: true               # enable auto memory extraction
      auto_extract: true          # LLM-based + keyword fallback
      extractor_model: ''         # empty = use cheapest model from your DSH config
      extractor_temp: 0.2
      confirm_threshold: 0.6      # memories below this confidence go to confirm queue
      interval_ms: 21600000       # 6h minimum between extractions

Self-Evolving Learning

- id: long-memory
  config:
    corrections:
      signal_words:               # trigger correction capture
        - '不对'
        - '错了'
        - '应该是'
      promote_threshold: 5        # minimum corrections before rule extraction
      rule_token_budget: 800      # max tokens for rule injection in context

Storage

    storage:
      path: '${DSH_HOME}/long-memory/long-memory.db'
      markdown_dir: '${DSH_HOME}/long-memory/markdown'

Tools

Tool	Purpose
mem_search	FTS5 + hybrid search across memories
mem_record	Persist a memory; auto-detects scope
mem_status	Storage + recall state
mem_stats	Aggregate statistics
mem_forget	Archive or delete; writes audit log
mem_confirm	Approve/reject queued sensitive memory
mem_scope_list	List all scopes
mem_scope_set_active	Set active scope filter

Web API

Method	Path	Purpose
GET	/plugins/dsh-long-memory/api/memories	List memories (scope/type/q filter)
PUT	/plugins/dsh-long-memory/api/memories/:id	Archive a memory (status=archived)
DELETE	/plugins/dsh-long-memory/api/memories	Delete a memory
GET	/plugins/dsh-long-memory/api/memories/stats	Per-scope memory counts
GET	/plugins/dsh-long-memory/api/confirm-queue	Pending sensitive memories
POST	/plugins/dsh-long-memory/api/confirm-queue	Approve/reject
GET/POST	/plugins/dsh-long-memory/api/config	Get/save plugin config
GET	/plugins/dsh-long-memory/api/corrections	List corrections (status/trigger filter)
POST	/plugins/dsh-long-memory/api/corrections/:id/extract	Promote correction to rule
POST	/plugins/dsh-long-memory/api/corrections/:id/ignore	Ignore correction
GET	/plugins/dsh-long-memory/api/rules	List rules (status filter)
POST	/plugins/dsh-long-memory/api/rules/:id/approve	Approve rule
POST	/plugins/dsh-long-memory/api/rules/:id/reject	Reject rule
POST	/plugins/dsh-long-memory/api/rules/:id/promote	Promote to AGENTS.md
GET	/plugins/dsh-long-memory/api/rules/:id/source	View source corrections
PUT	/plugins/dsh-long-memory/api/rules/:id	Edit rule
GET	/plugins/dsh-long-memory/api/stats	Aggregate stats

Architecture

┌──────────────────────────────────────────────────────────────────┐
│                    long-memory plugin                            │
├──────────────┬──────────────┬───────────────────────────────────┤
│  Memory &    │  Self-Evolving│  Browser UI (Sidebar + Content)  │
│  Recall      │  Learning     │                                   │
│              │               │  ┌───────────┬─────────────────┐ │
│ • memories   │ • corrections │  │  Sidebar   │  4 Tab Panel   │ │
│ • FTS5       │ • rules       │  │  • scope   │  • 教训         │ │
│ • embedding  │ • signal words│  │  • search  │  • 规则         │ │
│ • L7 extract │ • tool errors │  │  • stats   │  • 记忆         │ │
│ • scope      │ • rule inject │  │            │  • 画像         │ │
│ • KG         │ • decay       │  │            │  • 召回测试     │ │
│ • 8 tools    │ • conflicts   │  │            │  • 设置 Modal   │ │
│ • audit log  │ • AGENTS.md   │  │            │  • Toast/Popover│ │
│ • persona    │               │  │            │                 │ │
└──────────────┴──────────────┴───────────────────────────────────┘

Requirements

• Node ≥ 22.5 (uses built-in node:sqlite)
• DeepSeek Harness 0.1.0-rc.2+

Development

The host-side sources live in src/ as strict TypeScript; the browser bundle currently lives in src/client.js (pre-wrapped, re-sourced as TSX in a later pass). lib/ is build output only.

pnpm install
pnpm run typecheck   # tsc --noEmit
pnpm test            # vitest + legacy node scripts (migrations, tools e2e)
pnpm run build       # esbuild host/invariant bundles + per-module transform + client copy + tsc declarations
pnpm run check       # all of the above

The build keeps the flat per-module lib/*.js layout in sync with src/*.ts (the scripts/ test suite imports those module paths directly). Host DSH / cordis peer packages stay external — the DSH profile's node_modules provides them at runtime.

Migrations

#	File	Description
0001	0001_initial.sql	Core tables: memories, embeddings, audit_log, confirm_queue, schema_meta
0002	0002_l7_buffer.sql	L7 message buffer table
0003	0003_embedding_cache_key.sql	Composite PK for embeddings
0004	0004_corrections_rules.sql	Corrections + rules + usage_stats (self-evolving)

License

MIT — see LICENSE.