dsh-tdai-memory

GitHub: Scorp1o117/dsh-tdai-memory · npm: dsh-tdai-memory

Part of the DeepSeek Harness Enhancement Suite — Vision · Soul/Persona · Long-term Memory · Plugin Marketplace.

A port of TencentDB Agent Memory (Tencent Cloud's open-source four-layer memory system, originally an OpenClaw plugin) into DeepSeek Harness.

Features

• L0 conversation capture: every turn (turn end, request boundary) is written to raw conversation storage (JSONL + SQLite + FTS + vectors)
• L1 structured memory: a background pipeline uses an LLM to extract facts / preferences / events (persona / episodic / instruction) from conversations, stored in records/ + SQLite + FTS + vectors
• L2 scenes / L3 persona: scene blocks and user profile generation (pipeline-scheduled)
• Automatic recall injection: on every prompt assembly, relevant memories and the user profile are retrieved by the current user message and injected as dynamic context (the model "just remembers")
• Tools: tdai_memory_search (L1 structured search), tdai_conversation_search (L0 raw-text search)

The data directory reuses the existing ~/.memory-tencentdb/memory-tdai, so previously accumulated memories carry over seamlessly.

Architecture (porting approach)

Hard-won wiring details:

• Capture: session/flush listener (await semantics; must complete before headless exits); turn/start timestamps as the L0 cursor floor; turn-id dedup
• Headless one-shot runs: wait for core.handleSessionEnd() inside flush (L1 extraction finishes before exit; otherwise the 5s shutdown timeout kills it)
• Recall injection: must be registered on agent.ctx (assembly runs in the agent scope; root listeners never see it); attach one tick after session/created by resolving the agent from the agents service

Configuration (profile patch + settings)

Configuration is settings-namespace driven: the profile patch is the base layer, and the tdai-memory: section of $DSH_HOME/settings.yaml overrides it (LLM/embedding keys live in settings.yaml). The Web UI Settings → 记忆 section edits every field (v0.2.0, write-only keys); TdaiCore is built at startup, so changes apply after a restart.

# $DSH_HOME/settings.yaml
tdai-memory:
  llm:
    apiKey: 'sk-...'
  embedding:
    apiKey: 'local-no-key'

# profile patch (base layer)
- id: tdai-memory
  name: 'dsh-tdai-memory'
  config:
    extraction:
      enabled: true
      enableDedup: false      # dedup LLM output parsing is flaky; off by default
    llm:                      # L1/L2/L3 extraction model (OpenAI-compatible)
      baseUrl: 'https://opencode.ai/zen/go/v1'
      model: 'mimo-v2.5'      # deepseek-v4-flash produces invalid extraction JSON
    embedding:                # vectors (OpenAI-compatible /v1/embeddings)
      baseUrl: 'http://127.0.0.1:8088/v1'
      model: 'Qwen3-Embedding-0.6B'
      dimensions: 1024
      sendDimensions: false

Install

dsh plugin --profile web add dsh-tdai-memory

then mount it in $DSH_HOME/profiles/web/cordis.patch.yml:

- insert:
    - id: tdai-memory
      name: 'dsh-tdai-memory'
      config: {}          # keys can live in settings.yaml instead

and restart dsh web. LLM/embedding API keys can be set in the Web UI settings page (记忆 / Memory) or directly in settings.yaml under tdai-memory:.

Note for users

• This plugin is a standard profile bundle (dsh.bundle.patch): dsh plugin --profile web add dsh-tdai-memory installs and mounts it in one step — no manual cordis.patch.yml edits needed.
• The settings section needs the dsh-host-apiproxy namespace allowlist; the plugin patches it automatically on first start — restart dsh web once more and the section appears. A dsh update overwrites the patch; the next plugin start re-applies it.
• Settings changes apply after a restart (TdaiCore is built at startup).
• Tested against DSH 0.1.0-rc.6.

Known trade-offs

• Extraction model: mimo-v2.5 extracts correctly but takes 20-30s per call (background execution, does not block the conversation); deepseek-v4-flash is fast but its JSON output is non-compliant (extracts 0)
• dedup: LLM conflict-detection output parsing is unstable (once caused stored=0); off by default; enable only with a more reliable model
• L1 memory vectors: written with storage (8088 embedding is fast); L0 vectors run as a background task, drained by destroy() on headless exit
• Upgrades: after pulling new upstream code, rerun npx tsc -p dsh-tsconfig.json in the tdai project dir (output in dist-dsh/)

License

MIT