What it does

memory-lancedb-pro plugs into OpenClaw agents and stores long-term memory in LanceDB. It auto-captures preferences, decisions, and project context from conversations, then injects relevant memories back into future prompts. The plugin also indexes canonical files like MEMORY.md and session transcripts for semantic retrieval.

The interesting bit

The retrieval stack is over-engineered in a useful way: vector search plus BM25 full-text, fused and reranked by a cross-encoder. Memories also decay via a Weibull model—important stuff lingers, noise fades. It’s the kind of “boring” infrastructure that actually determines whether an agent feels coherent across sessions.

Key highlights

• Hybrid retrieval (vector + BM25) with cross-encoder reranking
• LLM-powered 6-category extraction: profiles, preferences, entities, events, cases, patterns
• Weibull decay for “intelligent forgetting”—no manual cleanup
• Multi-scope isolation: per-agent, per-user, per-project boundaries
• Full CLI toolkit: backup, export/import, migration, upgrade commands
• Community setup script handles install, broken configs, and path detection

Caveats

• Requires AVX CPU instructions; AVX2 needed on some Linux x64 builds or LanceDB native search may SIGILL (fallback to JS cosine ranking available)
• Currently in beta (v1.1.0-beta.10); adapted for OpenClaw 2026.3+ hook architecture, breaking from older before_agent_start hooks
• npm install requires absolute path in plugins.load.paths—the README calls this “the most common setup issue”

Verdict

Worth it if you’re running OpenClaw agents and tired of repeating yourself every session. Skip if you’re not on OpenClaw or your deployment target lacks AVX support.