One-Sentence Positioning

Port Tencent Cloud's four-layer memory system (TencentDB Agent Memory, formerly OpenClaw plugin) into DeepSeek Harness: automatically save every conversation round, use LLM to extract facts/preferences/profiles, and automatically inject relevant memories into the prompt based on the current message when asking questions again—so the model can "recall" past events out of thin air.

Core Capabilities

• Automatically save every conversation round to local SQLite + JSONL + FTS + vector index (L0 raw capture)
• Background LLM extracts facts, preferences, and events from conversations, writing to structured memory (L1) + auto-generating user profiles and scene blocks (L2/L3)
• Before each message is sent, automatically retrieve relevant memories based on current user input and inject as dynamic context into the prompt (no manual search needed)
• Model can actively call tdai_memory_search (structured memory search) and tdai_conversation_search (raw conversation search) two tools
• Web UI settings page has a dedicated "Memory" section for visual editing of all configurations (data directory, extraction model, Embedding, toggles), leaving key empty preserves original value
• Data directory reuses ~/.memory-tencentdb/memory-tdai, Tencent Cloud version legacy memories can be directly migrated for use

Technical Implementation

• Language: JavaScript (ESM, Node native, TypeScript only used for vendored core)
• Key Dependencies: @deepseek-ai/schemastery (config schema), @deepseek-ai/dsh-tools (tool registration), @tencentdb-agent-memory/tcvdb-text (vector DB binding), sqlite-vec (local vector search), @node-rs/jieba (Chinese word segmentation)
• Architecture Pattern: Cordis plugin (profile bundle). index.js listens to session/event+session/flush for capture, listens to system-prompt/assemble for recall injection; client.js renders Web UI settings section; recall-inject.js serves as preset line-level fallback recall injector. Underlying core TdaiCore comes from vendored Tencent Cloud version, zero changes via StandaloneHostAdapter directly connects to OpenAI-compatible LLM/Embedding
• Entry Files: index.js (host-side Cordis plugin) + client.js (browser-side ModuleLoader plugin) + recall-inject.js (fallback line in agent preset)

Use Cases

For users who use the same conversation/Agent in DSH for a long time and want the model to remember facts you've mentioned (project background, personal preferences, historical decisions); or those who want to "ask one question and recall what was discussed last time" instead of restating the background every time. This plugin is not needed for one-off Q&A scenarios—you need a long-running session/Agent to notice its value.

Prerequisites and Compatibility

Installation

dsh plugin --profile web add github:Scorp1o117/dsh-tdai-memory

Configuration Options

FAQ

Q: Do I need to manually modify cordis.patch.yml after the installation command finishes?

A: No. The - insert declaration in cordis.patch.yml has already registered tdai-memory as a profile bundle; dsh plugin automatically merges and mounts it during package installation.

Q: Why don't changes to the settings page take effect?

A: TdaiCore is built once at plugin startup; runtime changes to settings don't rebuild it. You must restart dsh web after modifying for changes to actually take effect (there's also a warn in index.js:380).

Q: Can I use deepseek-v4-flash as the extraction model?

A: Not recommended. Its JSON output is often non-compliant, which can cause L1 extraction to yield zero items; README defaults to mimo-v2.5, which takes 20-30s per run but executes in the background without blocking conversations.

Q: Do I have to install node-llama-cpp?

A: No, it's not required. It's only an optional peer for the "fully local embedding" backend; using OpenAI-compatible remote embedding (default) doesn't need it.

Q: Will memories accumulated with the Tencent Cloud version on old machines be lost?

A: No. The data directory is fixed at ~/.memory-tencentdb/memory-tdai; content written by the Tencent Cloud version is preserved as-is, and the new plugin reuses it directly.

Q: Can the model actively search old memories?

A: Yes. After registration, tdai_memory_search (L1 structured search) and tdai_conversation_search (L0 raw search) will appear in the tools list for the model to actively call.

Q: Will data remain after uninstalling the plugin?

A: Yes. dsh plugin remove only uninstalls the package and unmounts it; ~/.memory-tencentdb/memory-tdai won't be cleared; to completely delete, manually rm that directory.

Difficulty Level

Advanced — requires you to select and integrate an OpenAI-compatible extraction LLM and embedding service yourself (the default http://127.0.0.1:8088 Qwen3-Embedding may not exist), and understand the four-layer memory architecture's L0/L1/L2/L3 to tune parameters properly; it runs out of the box, but to extract value you need to choose the right model and embedding backend.

Known Issues and Limitations

• Extraction model deepseek-v4-flash outputs non-compliant JSON, which can cause L1 to extract zero items; mimo-v2.5 is recommended by default (20-30s/run, executes in background) (README.md:125-127)
• Conflict detection (dedup) LLM output parsing is unstable, has caused stored=0, off by default; enabling requires a more reliable model (README.md:128-129 / index.js:63)
• Single mimo-v2.5 extraction takes 20-30s; with multiple rounds in long sessions, overall L1 extraction time is considerable (README.md:125)
• L0 vector writes go through background tasks; in headless one-shot tasks, destroy() drains them on exit, and excessive task volume may lose tail items (README.md:130-131)
• The seed-runtime.js in vendored core has a FIXME: teardown waits for L1 to finish then destroys, but L2/L3 pipelines may be interrupted before completion (vendor/tdai/core/seed/seed-runtime.js:8-14)
• The tcvdb.js in vendored core has a TODO: vector collection creation is not delayed to first use, which may block plugin initialization during startup (vendor/tdai/core/store/tcvdb.js:178)
• Web settings page configuration changes require restarting dsh web to take effect; runtime settings changes only write to disk without rebuilding the core (README.md:111 / index.js:379-382)
• After upgrading if pulling new upstream code, need to re-run npx tsc -p dsh-tsconfig.json in the tdai project directory (README.md:132-133)

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, 0.1.0-rc.7, and 0.1.0-rc.8.

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