memory-lancedb-pro is a long-term memory plugin for the OpenClaw agent gateway, extending the built-in memory-lancedb plugin with substantially more retrieval machinery. Where the built-in plugin offers basic vector search, this one adds BM25 full-text search, hybrid fusion of the two, cross-encoder reranking, recency and time-decay weighting, length normalization, MMR diversity, noise filtering, adaptive retrieval, multi-scope isolation, and a management CLI.
Retrieval runs a pipeline: vector (LanceDB ANN) and BM25 results are fused (RRF with a BM25 boost), optionally reranked by a cross-encoder (Jina, SiliconFlow, Voyage, Pinecone, or vLLM endpoints, with graceful fallback), then passed through a multi-stage scoring stack (recency boost, importance weighting, length normalization, time decay, hard minimum score, MMR de-duplication). It works with any OpenAI-compatible embedding provider (Jina, OpenAI, Gemini, Ollama) and handles long documents via sentence-boundary chunking with averaged embeddings.
Memory is partitioned into scopes (global, agent, project, user, custom) with per-agent access control, so agents don't pollute each other's memory. Lifecycle hooks provide auto-capture (extracting preferences, facts, decisions, and entities at conversation end) and auto-recall (injecting relevant memories into the prompt), both gated by adaptive logic that skips greetings and trivial turns.
It exposes agent tools (memorystore, memoryrecall, memoryforget, memoryupdate, memorylist, memorystats) plus an optional self-improvement governance layer that logs learnings and errors to .learnings/ files and can scaffold reusable skills. A memoryReflection mode supports reflection-based rule inheritance, and a markdown-mirror option writes a readable copy of memories alongside the LanceDB store. An optional pipeline distills OpenClaw session transcripts into high-signal atomic memories rather than dumping raw logs.
