CortexReach/memory-lancedb-pro
Memory Lancedb Pro
memory-lancedb-pro is a plugin that gives AI agents persistent long-term memory stored in LanceDB. It saves and recalls facts, preferences, decisions, and entities across conversations, capturing new memories after each turn and injecting relevant ones into the agent's context before each reply, with no manual save command needed.
What sets it apart is how memories are extracted and retrieved. A language model classifies conversation content into profile, preference, entity, event, case, and pattern categories, then deduplicates and merges results before storage. Recall blends vector similarity with BM25 keyword matching and can reorder results with a cross-encoder reranker. Memories are also separated into global, agent, user, project, and custom scopes, with per-agent rules that block reads or writes to scopes an agent is not permitted to use.
Beyond that core, the plugin scores memories by recency, access frequency, and importance, so stale entries fade while frequently recalled ones move between core, working, and peripheral tiers. It filters out low-quality content such as refusals and greetings, saves a session summary when a new chat starts, and can index memory files and transcripts so search returns file paths, line spans, and citations. Embeddings can come from any OpenAI-compatible provider, and concurrent writers are coordinated with a file lock, optionally upgraded to a Redis lock when several machines share a store.
A command-line tool lists, searches, exports, imports, re-embeds, upgrades, and migrates stored memories, and a setup script handles installing, upgrading, and repairing the plugin in one command.
