Cognee: Open-Source AI Memory Platform for Agents

What is Cognee?

Cognee is an open-source AI memory platform designed for agents to remember across sessions, ingesting documents to extract and store entities and relationships as a searchable knowledge graph. Backed by vector embeddings, it uses four API calls (remember, recall, forget, improve) for information storage, querying, deletion, and memory tuning. The platform runs locally, in Docker, or as a managed cloud service, and ships official clients for Python, Rust, and TypeScript.

Core Capabilities for Agent Memory

• ✓Four-operation API: remember, recall, forget, and improve for full memory lifecycle management
• ✓Hybrid storage combining vector embeddings with graph relationships for meaning-based and structure-based retrieval
• ✓Session memory with background sync to the permanent knowledge graph
• ✓Claude Code plugin that captures prompts, tool traces, and responses into session memory and persists context across context resets
• ✓Multi-backend architecture: LanceDB, Postgres/pgvector, Neo4j, Qdrant, ChromaDB, Weaviate, Milvus, and more via community adapters
• ✓Multi-tenant isolation support with per-user, per-dataset graph and vector databases on compatible backends
• ✓Docker images for the API server and MCP server published on Docker Hub on every push to main
• ✓Benchmark results on the BEAM long-context benchmark: 0.79 at 100K tokens, 0.67 at 10M tokens

Real-World Applications of Cognee

• •Building customer support agents that track interaction history, failed actions, and resolved cases across sessions
• •Creating SQL copilots that distill and reuse expert-level query patterns and reasoning
• •Unifying data from multiple company channels into a single knowledge graph that any agent can query
• •Giving Claude Code persistent memory that survives context resets and session boundaries
• •Running an AI research assistant that maintains a growing ontology of papers, concepts, and their relationships
• •Building domain-specific expert systems where agents learn from feedback and improve over time

Getting Started with Cognee

Install via pip, uv, or poetry. Python 3.10 to 3.14 is required. Run `uv pip install cognee` for the core package, or add `[postgres]` for the Postgres stack. Configuration is done through environment variables: set `LLM_API_KEY` for the language model, then point at your chosen backends with variables like `DB_PROVIDER`, `VECTOR_DB_PROVIDER`, and `GRAPH_DATABASE_PROVIDER`. A `.env.template` in the repo shows all available options. Docker users can pull `cognee/cognee:main` directly from Docker Hub or use the provided docker-compose file with optional profiles for the UI, MCP server, Postgres, and Neo4j.

Integrating Cognee into Your AI Agents

Import cognee in your async Python code and call the four operations. `await cognee.remember("your text here")` stores permanently in the knowledge graph. Pass `session_id` to route through session cache instead. `await cognee.recall("your query")` queries with automatic routing to the best retrieval strategy. `await cognee.forget(dataset="name")` deletes a dataset. Environment flags let you trade off memory quality for speed: `AUTO_FEEDBACK=false` disables the post-query self-tuning step, and `CACHING=false` disables session memory entirely. The CLI offers the same operations with `cognee-cli remember`, `cognee-cli recall`, and `cognee-cli forget --all`.

Exploring AI Agent Memory Solutions

Common Questions About Cognee

The problem it solves

Agent frameworks give you tools to think and act, but nothing remembers what happened last week, last month, or last quarter. Vector stores solve the retrieval problem but lose the relationships between facts. Rag systems need to be rebuilt for every session. Developers end up hand-rolling brittle context windows, losing the long-term learning that makes an agent actually useful over time.

Who should try it — and who should skip

Reach for Cognee if you are building a multi-session AI agent and you do not want to assemble a memory stack from a vector store, a graph database, and a session cache. It is particularly useful for Python or TypeScript teams that want a self-hosted option with a research-backed architecture. The Postgres-integration makes it practical for solo developers and small teams who want a single-service memory layer during development. Skip it if your agents do not need persistent cross-session memory, if you are already invested in a different agent framework with its own memory system, or if you need production-grade graph isolation today and cannot wait for the managed offering.

Related repositories