# remembrane

A local-first memory library for AI agents that stores everything in one SQLite file with no required dependencies, combining exact vector and BM25 recall, conflict detection and a time-travel history, with LangChain, CrewAI and MCP integrations.

remembrane gives an agent persistent memory across sessions without a cloud API, a vector database or a heavyweight framework: the entire memory lives in a single SQLite file that can be copied, backed up, diffed or deleted, and the default embedder is pure standard library, so installing the package pulls in nothing else.

Recall results are ranked by a weighted sum of similarity, recency decay, importance and outcome-earned usefulness, and recalled memories are reinforced, similar to spaced repetition. It scores every memory with exact hybrid vector-plus-BM25 search rather than approximate nearest-neighbor search, so results are guaranteed complete, and it supports conflict-aware recall that admits uncertainty, a time-travel journal over memory history, and budget-capped context packing that is exactly optimal when numpy is present, with deterministic behavior that can be unit-tested in CI.

Built with Python, installed with pip; ships adapters for LangChain and CrewAI plus a built-in MCP server. MIT licensed.

Website: https://github.com/satyasairay/remembrane