Give agents persistent memory across sessions with a free open-source platform
This tool solves the problem of agents forgetting previous content after a conversation or task ends. It turns documents, code, and chat logs into a single knowledge graph that agents can search and reuse at any time. It runs locally for free, and basic features work without an API key.
Recent updates: local model support and Google integration
The recent v1.6.1 release includes Google Drive and Gmail OAuth connection features in the SDK. It also adds the ability to stream visualization data in chunks when processing large graphs. v1.6.0 introduced a workflow for building and searching text memory using only local models, without cloud LLM keys. If no LLM is configured, LLM-dependent improvement steps are automatically skipped, and pipeline recovery preserves completed documents after a crash.
Memory construction and retrieval methods
Cognee converts text into entities, relationships, and searchable chunks, and creates code as symbol and dependency graphs. Content learned in a session is refined and stored in permanent memory. During retrieval, it selects and fetches relevant items from the graph, vector, or code context. You use remember to store content, recall to retrieve context and answers, improve to enrich memory, and forget to delete specific items.
Connecting agents and applications
You can install Claude Code and Codex plugins to connect memory to existing agents. It is available in various environments through MCP clients, Python, TypeScript, Rust applications, and the HTTP API. To check the local installation status in the UI, use the cognee-cli -ui command. The UI launcher requires Node.js/npm, and the MCP service requires Docker.
Pre-use considerations
Python 3.10 to 3.14 is required. If running without an LLM key, the GLiNER model and local embedding models are downloaded on first use. Additional LLM configuration is needed for generated answers or media processing. BEAM evaluation reported scores of 0.79 for 100K token context and 0.67 for 10M tokens, but these are results under specific data formats and retrieval settings, so you should verify the methodology before comparing with other systems. Since the Postgres-based graph storage feature is in the demo stage, consider a licensed product for production use.
