This Open-Source Tool Gives AI Agents Real Memory — Running on Ollama

Your AI agent has amnesia. Every session starts from zero — yesterday's context, decisions, and facts, gone. Cognee is an open-source AI memory platform that fixes this by turning your documents into a living knowledge graph — and it runs 100% locally on Ollama, no cloud, no API key.

RAG vs a real knowledge graph

RAG is a filing cabinet: chunk documents, embed them, retrieve the closest chunks. It finds similar text, but it can't connect facts across documents — ask something that requires joining fact A in one file with fact B in another and plain RAG shrugs.

Cognee builds a knowledge graph instead: entities and relationships extracted from your data, linked into a structure the agent can actually traverse. That's the difference between "here are five paragraphs that mention X" and "X works at Y, which acquired Z."

How it works: the ECL pipeline

Cognee's pipeline is Extract → Cognify → Load: pull content from your sources, distill entities and relations into the graph, and serve it back as queryable memory. It maintains two kinds of memory — the semantic graph plus vector embeddings — so you get graph reasoning and similarity search from one store.

The local demo

On my RTX 4060, all local:

• LLM: llama3.1:8b on Ollama
• Embeddings: nomic-embed-text
• Add facts → watch the graph build → visualize it → ask questions that require reasoning across documents

The cross-document answers are the money shot: the graph connects facts no single chunk contains.

Pros and cons

Pros

• Local-first: your data never leaves the machine
• Pluggable: swap LLMs, embedders, and stores freely
• One Postgres can hold everything — no exotic infra
• BEAM and a growing ecosystem around it

Cons

• Graph building costs tokens and time on big corpora — an 8B local model takes a while
• Extraction quality tracks the LLM you give it
• Young project; APIs still moving

Who it's for

Anyone building agents that live longer than one session: support bots that remember customers, research assistants that accumulate a corpus, personal agents that know your notes. If your agent needs memory and your data can't leave the building, this is the strongest local-first option I've tested.

Repo: github.com/topoteretes/cognee · Docs: docs.cognee.ai — full local run in the video.