CortexReach/memory-lancedb-pro

memory‑lancedb‑pro – Long‑term memory plugin for OpenClaw agents

What it is – A production‑grade OpenClaw plugin that gives an AI agent a persistent “brain”. It stores conversation snippets, preferences, decisions and other context in a LanceDB vector store and automatically recalls the most relevant pieces when the agent replies.

Key capabilities

Feature	How it works
Auto‑capture	Every chat turn is sent to the plugin; no manual memory_store calls are needed.
Smart extraction	An LLM classifies each captured chunk into six categories (profiles, preferences, entities, events, cases, patterns).
Hybrid retrieval	Queries are embedded, then searched with both ANN vector similarity and BM25 full‑text search. Results are fused and optionally reranked by a cross‑encoder.
Intelligent forgetting	A Weibull decay model reduces the weight of old or low‑importance memories, letting noise fade.
Context injection	Before the agent builds its reply, the plugin injects the top‑ranked memories into the prompt automatically.
Scope isolation	Memories are partitioned per‑agent, per‑user and per‑project, preventing cross‑talk leakage.
Provider‑agnostic embeddings	Works with OpenAI, Jina, Gemini, Ollama or any OpenAI‑compatible API.
Tooling	CLI for listing, searching, exporting, importing, re‑embedding, upgrading and migrating the database.
Dreaming support	Optional side‑car that treats “dream” reports as public artifacts and indexes them.

Installation

1. One‑click script – curl …setup-memory.sh && bash setup-memory.sh (handles install, config, and service restart).
2. OpenClaw CLI – openclaw plugins install memory-lancedb-pro@beta.
3. npm – npm i memory-lancedb-pro@beta (then add the absolute plugin path to plugins.load.paths).

Configuration snapshot (the README’s default JSON):

{
  "plugins": {
    "slots": { "memory": "memory-lancedb-pro" },
    "entries": {
      "memory-lancedb-pro": {
        "enabled": true,
        "config": {
          "embedding": {
            "provider": "openai-compatible",
            "apiKey": "${OPENAI_API_KEY}",
            "model": "text-embedding-3-small"
          },
          "autoCapture": true,
          "autoRecall": true,
          "smartExtraction": true,
          "canonicalCorpus": { "enabled": true, "syncOnSearch": true },
          "dreaming": { "enabled": false },
          "extractMinMessages": 2,
          "extractMaxChars": 8000,
          "sessionMemory": { "enabled": false }
        }
      }
    }
  }
}

Why these defaults? – They turn on hands‑free learning (autoCapture + smartExtraction) and automatic recall (autoRecall). The low extractMinMessages makes extraction trigger after a normal two‑turn exchange, while disabling per‑session memory avoids polluting the long‑term store with transient summaries.

Runtime architecture (as described in the README):

• index.ts registers the plugin with OpenClaw and wires lifecycle hooks (before_prompt_build).
• store.ts talks to LanceDB (vector + BM25 indexes, CRUD operations).
• embedder.ts abstracts the embedding provider.
• retriever.ts performs hybrid search, fusion, cross‑encoder reranking and applies the decay boost.
• smart-extractor.ts runs the LLM classification.
• tools.ts exposes the agent‑side tools (memory_recall, memory_store, memory_forget, memory_update, plus optional management tools).

Typical workflow

1. Capture – After each assistant reply, the plugin extracts salient facts and stores them as rows in LanceDB, also writing the original markdown files (MEMORY.md, memory/**/*.md, etc.) for human reference.
2. Recall – When the next user query arrives, the before_prompt_build hook runs: the query is embedded, hybrid‑searched, reranked, and the top hits are inserted into the prompt as context snippets.
3. Decay – Each stored row carries a timestamp and importance score; the Weibull decay engine gradually lowers its relevance unless it is accessed frequently.

Hardware note – LanceDB’s native cosine ANN requires AVX/AVX2. On CPUs lacking those instructions you can disable native cosine (retrieval.disableNativeCosine: true or env var MEMORY_LANCEDB_DISABLE_NATIVE_COSINE=1).

Ecosystem helpers

• Setup script – Handles fresh install, upgrades, config repair, and un‑install.
• Skill package – memory-lancedb-pro-skill lets a Claude‑Code or OpenClaw agent configure the plugin via natural‑language commands.
• Video tutorials – YouTube and Bilibili walkthroughs cover installation, hybrid retrieval internals, and debugging.

Who would use this?

• Developers building long‑running OpenClaw assistants that need to remember user preferences, past decisions, or project context across sessions.
• Teams that want a “brain” for their AI agents without building their own vector store or forgetting logic.
• Anyone who prefers a single, configurable plugin that works with any OpenAI‑compatible embedding provider.

---

Bottom line – memory‑lancedb‑pro is a fully‑featured, production‑ready memory layer for OpenClaw agents, combining semantic vector search, classic keyword search, decay‑based forgetting, and LLM‑driven extraction, all wrapped in an easy‑install plugin.