A use case we now serve did not exist as a category ten months ago: agents that make decisions on top of several data sources, improve on their own results and reconcile their memories in the loop.
That only works when the agents share what they know. In a five-minute video, Vasilije runs Claude Code and Codex in separate sessions and separate repositories on one company brain, and cognee 1.6.1 now brings Google Drive and Gmail into that memory.
Further down: that video, a guide to what really has to run on your own hardware for agent memory to count as local, everything that shipped, and where to find us through November.
Vasilije was in San Francisco for a hackathon with Bright Data, a working session on memory protocols, and a company brain talk at Docker the week they launched Cloud Sandboxes. If you are in town for SF Tech Week, come find us at the Scalekit hackathon on the 7th, or just reply to this email.
Claude Code renames a feature in one repository and records why. Codex, in a separate session and a different repository, reaches the same capability, sees that the name moved and sees the reason. Nothing was copied between them: both were reading the same company brain.
Moving an agent onto Ollama or llama.cpp does not make its memory local, because extraction, embeddings and reranking can each still call a hosted model. Xavier Francuski’s guide splits locality into four levels and ends with a test anyone can run: preload the models, take the machine offline, and check that a fresh session can still recall, revise and retrieve a memory.
Eighty-five pull requests merged in the week cognee 1.6.0 shipped, and 1.6.1 followed on 24 September.
Google Drive and Gmail, synced. 1.6.1 adds an OAuth connector for Google Drive, shared drives included, plus Gmail ingestion, with sync progress shown in the UI.
A memory plugin for Hermes. Nous Research’s Hermes agents can now load cognee as their memory provider, a plugin users kept asking for.
Web pages straight into memory. Keenable is now a built-in fetch backend: set one environment variable, pass a URL to remember(), and the page lands in the graph as clean Markdown.
Big graphs that open. The visualization endpoint streams large subgraphs in chunks instead of one payload, so a big memory graph renders without exhausting the server.
Safer self-hosting defaults. Random secrets replace the hardcoded token secret and default-user password, and cognee-mcp adds X-Api-Key auth for self-hosted mode.
A Node.js agent that remembers. In a new six-minute tutorial, Brandon Minnick runs @cognee/cognee-ts, an embedded Rust engine inside the Node process that keeps the graph on disk next to the app.
Four places to find us between now and November, in San Francisco, Menlo Park and London.
Not near any of those? If you are at a company that wants to host one with us, a meetup, a workshop or a hackathon, please get in touch.
Forward this issue to a builder who cares about self-improving AI memory.
Know someone using agents?
Forward this issue to a builder who cares about self-improving AI memory.