The Cognee Blog

A reference AI agent memory architecture with diagram: working, episodic, semantic and procedural memory, the write and read paths, and storage trade-offs.

How ChatGPT memory works (saved memories and chat history reference), what it cannot do, why the API has no memory, and how to add persistent memory to your own app with cognee.

How knowledge graphs and LLMs complement each other: LLMs build the graph, the graph grounds the LLM, GraphRAG connects multi-hop context, and the graph becomes agent memory.

What a temporal knowledge graph is, valid time vs transaction time, how contradictions are resolved with history kept, and why AI agent memory depends on it.

What has to run on your own hardware for agent memory to count as local: four levels of locality, where cloud calls hide in extraction, embeddings, and reranking, how much hardware each stage needs, and a disconnect test for fully offline memory.

Compare LangMem, Mem0, cognee, Zep, and Letta by how much of the stack each AI memory tool takes over — from memory primitives over your own store to managed temporal graphs and memory built into the agent runtime.

Compare how Mem0, Zep, cognee, Letta, and LangMem let you evaluate AI memory — open benchmark harnesses, retrieval-versus-answer scoring, published results, and what you still have to test yourself.

Compare Graphiti, Mem0, cognee, Basic Memory, and Letta for small models and local agents — which AI memory tools need generative inference, what runs offline, and how much work lands on the agent’s own model.

Compare DeepEval, Ragas, Promptfoo, Inspect AI, and Opik across output, retrieval, trajectory, and task-outcome evaluation layers — what each tool actually tests and how far it goes beyond scoring a single model response.