Cognee Review: Open-Source AI Memory for Agents

Key highlights:

• ECL pipeline — Extract → Cognify → Load. Documents become a graph of entities and relationships, persisted to Neo4j, Kuzu, or PGVector, then exposed through a tiny memory API.

• Four-verb API — `remember`, `recall`, `forget`, `improve`. That's the whole surface for storing, querying, deleting, and refining memory.

• Runs locally — pip/uv install or Docker Compose, self-hosted graph + vector storage, works with any LLM through a Python SDK or OpenAI-compatible API.

• Session + permanent memory — a fast session cache that syncs into the permanent graph in the background.

• MCP server + plugins — first-class Model Context Protocol server, plus Claude Code and OpenClaw plugins for drop-in agent memory.

The Problem Cognee Solves

Every stateless agent has the same amnesia. A code-review assistant that spent a month learning 14 recurring bug patterns in your repo starts its next review knowing nothing. A support agent that learned a customer's account quirks forgets them the moment the session ends. The industry's default answer — retrieval-augmented generation over a vector store — helps, but it has a structural ceiling: vector search finds text that's semantically near your query, but it doesn't know how facts relate.

Ask a vector store "which of our customers were affected by the outage that the billing bug caused?" and it will happily return chunks mentioning "outage," "billing," and "customers" — without ever connecting this bug to that outage to those customers. Relationship reasoning is exactly where pure similarity retrieval falls down, and it's the gap Cognee targets.

Cognee's bet is that long-context, multi-hop questions need a graph. Graphs are slower, more complex, and more expensive than vectors — the Cognee team knows this and picked the graph anyway, because relationship reasoning is a hard requirement they don't think embeddings alone can satisfy.

How Cognee Works: The ECL Pipeline

Where a normal RAG stack is "chunk → embed → retrieve," Cognee runs an ECL pipeline:

1. Extract — ingest data in any format (docs, chats, code, APIs) and pull out entities and relationships.

2. Cognify — structure those entities against a generated, cognitive-science-grounded ontology, so the graph has meaningful relationship types rather than raw co-occurrence.

Running It Self-Hosted with Docker

For anything beyond a script, run the server. Cognee publishes prebuilt images on every push to main, so you don't even need to clone:

docker run --env-file ./.env -p 8000:8000 --rm -it cognee/cognee:main

Or build from source with Compose and opt into the pieces you want:

Read the Full Deep Dive

This is a condensed version. The full article includes architecture diagrams, code examples, comparison tables, and practical use cases.

👉 Read the complete analysis on andrew.ooo

Topics: cognee, ai-memory, knowledge-graph, ai-agents, rag