brainctl

Persistent agent memory is provided by a SQLite-backed brain that stores, retrieves, and decays memories across sessions. You initialize a local brain.db file, then use an agent_id to log events and query historical context via CLI, Python API, or MCP server.

Cross-agent collaboration is enabled by sharing one brain.db file across multiple agents on a single machine. Agents coordinate through handoffs, triggers, and trust dynamics, reading and writing to the same SQLite-backed knowledge graph simultaneously.

Building a knowledge graph involves using CRUD operations to create entities and edges within a SQLite-backed brain. You store entities, define relationships as edges, and record decisions, allowing agents to retrieve structured reasoning context through FTS5 or vector search.

Python 3.11+ and SQLite are required for local agent memory, while sqlite-vec is an optional dependency for vector search and semantic recall. FTS5 full-text search is available by default for cross-table retrieval across memories, events, and entities without vector extras.

An MCP server surface exposes 201 tools for programmatic control of agent memory via remote agents. By installing the optional MCP extras, you can initialize brain.db and manage memories, events, and entities directly through MCP endpoints.

Single-machine persistence is fully supported without a network connection because brainctl uses a local SQLite database as its spine. All memory lifecycle operations, including storage, retrieval, and decay, function offline within the local brain.db file.