Agents that remember: cognee 1.0 is out

cognee 1.0 shipped on June 26. It's open-source memory for agents - the part that comes after retrieval, where the agent needs to remember what it did last session instead of starting cold every time.

cognee is a dltHub partner, and uses dlt under the hood for structured data ingestion.

What's in the release

• Runs on a single Postgres. No separate vector store and graph store to provision, monitor, and keep in sync.
• Rust core, for faster cold starts.
• Self-improving memory loop - carries forward what's used, lets the rest decay.
• TypeScript SDK for Node agents.
• Managed Cloud, if you'd rather not run the Postgres yourself.

Why memory you can read matters

A vector store gives you similarity you can't audit. A graph you can traverse, inspect, and correct when an edge goes stale - and in a long-running agent, something always goes stale. Because the core is open source, the memory is yours to look at, not a vendor's opaque internal format.

What's good about it

• Data in any format, one memory layer. Docs, chats, APIs, warehouses - ingest it, and every agent can recall it. You're not maintaining a separate pipeline per source.
• It improves with use. Feedback feeds back into the graph, so the memory sharpens over time instead of going stale the moment you load it.
• A clean API. Four operations - remember, recall, forget, improve. That's the whole surface.
• Runs local or on-device. Open source at the core, with the new Rust engine aimed at edge cases where latency and privacy matter. Start on your laptop, scale to Cloud when you need to.
• Multi-tenant isolation. Per-user and per-group graphs with read/write/delete/share permissions, not just namespace separation - which matters the moment you ship memory to more than one customer.

They've also published benchmarks against Mem0, Graphiti, and LightRAG with the eval code open, which is the right way to make a claim.

What's new about Cognee 1.0?

• Continuously self-improving memory. As your agents interact with your memory, it self-improves and takes only what matters forward
• Cognee Cloud, our managed memory SaaS. Move from local to production in one line, with no infrastructure to run.
• Users can attach Cognee to all popular agents via pip install cognee-mcp and a config file. Claude Code, Cursor, Codex, OpenClaw, Cline, and more.
• Cognee can now run on a single Postgres. No separate vector or graph store to operate.
• Cognee’s new Rust core offers on-device, sub-350ms memory cold start
• Developers now get a feature-rich TypeScript SDK for Node agents.

Try it

Read more on cognee.ai

Connects to Claude Code, Cursor, Codex, and the rest via a config file.

Congrats to Vasilije, Veljko, and the cognee team on 1.0.