Subconscious

Gives Claude a local semantic memory layer backed by ChromaDB and sentence-transformers. You get four tools: recall(task) searches past work by similarity, remember(task, answer) stores new pairs, forget(entry_id) deletes an entry, and stats() shows hit rates. The default embedding model is all-MiniLM-L6-v2 running on CPU, so nothing leaves your machine. Useful when you want an agent to skip redundant work by checking if it's already solved a paraphrased version of the same task. First run downloads 80MB for the model, then recall latency drops to milliseconds. Configurable threshold and TTL per entry. All state lives in ~/.subconscious-mcp/data as a persistent ChromaDB collection.