topoteretes/cognee

What it solves

Cognee provides AI agents with persistent, long-term memory across sessions, solving the problem of agents forgetting context or losing track of user preferences and domain knowledge over time. It allows agents to build a "company brain" by unifying data from various sources into a searchable and connected knowledge base.

How it works

Cognee creates a self-hosted knowledge graph that combines vector embeddings for semantic search and graph reasoning for relationship tracking. It uses cognitive-science-grounded ontology generation to organize data. The system supports a tiered memory approach: session memory for fast caching and a permanent knowledge graph for long-term storage. It can be deployed as a standalone API server, a cloud-managed service, or integrated via plugins (e.g., for Claude Code).

Who it’s for

Developers building AI agents that require long-term memory, domain-specific knowledge distillation, and the ability to share knowledge across multiple agents or tenants.

Highlights

• Unified Memory Layer: Can run the entire stack (relationships, embeddings, sessions, and metadata) on a single Postgres instance using pgvector.
• Hybrid Search: Combines vector and graph search to allow agents to recall information based on both meaning and structure.
• Agentic Integration: Includes dedicated plugins for tools like Claude Code to capture prompts and tool traces automatically.
• Multi-language Support: Provides official clients for Python, Rust, and TypeScript.
• High Performance: Outperforms previous state-of-the-art benchmarks on the BEAM long-context benchmark for agent memory.