Last time we promised you a session on who gets to see what once several agents and teams share one memory. Vasilije ran it on 4 November, and if you couldn't make it, the recording is right below. In the meantime we also shipped cognee 1.6.7, which keeps your memory safer and your searches sharper.
It's been a busy two weeks on the road. Veljko flew to New York to speak at AI DevCon in Brooklyn, then spent a full day hacking with the LOCALHOST crowd in Williamsburg. Back home in Berlin, Nikola took the stage at LeadDev to debate whether AI is making our startup scene better or worse.
If you're in San Francisco on 20 November or in Berlin on 26 November, come and say hi. We'd love to meet you in person.
Memory for a single agent is simple. Once several agents, teams or customers share it, someone has to decide who sees what. In the recording, Vasilije walks through how cognee handles that access and which boundaries he would set on day one.
cognee 1.6.7. Out since 6 November. Here's what changes for you, with all the details in the release notes.
Safer ingestion by default. cognee now refuses raw database connection strings and server file paths sent over HTTP unless you opt in. If you rely on those, take a quick look at the release notes before you upgrade.
Scoped search stays in scope. Ask for one document type and that's exactly what you get back, and quoted, scoped queries no longer need an LLM call.
Faster answers. PGVector, Turso and LanceDB now score a whole batch of queries in one go, and forgetting sessions is batched too.
At AI DevCon in Brooklyn, Veljko gave the talk “Memory Without Deterministic Coordinates Is Just Hearsay”. If you weren't in the room, you can now watch the whole talk on YouTube.
On 10 November Nikola Zivkovic, our VP of Engineering, joined the panel at LeadDev Berlin for The Big Debate: is AI changing Berlin's startup culture for the better? Together with Sebastiano Castellani and Raz Schweiger-Shuty, moderated by Lucy Adams, they got into whether AI favours startups or scale-ups, what an exit looks like when half the codebase was generated, and what people still bring to the table.

On 7 November we co-hosted LOCALHOST with dmarz at The School in Williamsburg, a full day of hacking on open weights, open harnesses, open memory and private AI. Thank you to everyone who came and built with us.

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.