An agent that knows a fact but not when it was true will hand you last quarter’s answer with full confidence. So over the past two weeks we worked on improving cognee’s understanding of time and conflict resolution: cognee now confidently answers questions like “what changed since Tuesday?”, and if two of your sources disagree, it shows you the conflict instead of quietly picking a side.
I am still in San Francisco, where I hosted Mem Off, an open memory showdown, together with our friends from HelixDB and claude-mem. Everyone brought their own memory setup and ran it on the same scenarios: facts that go stale, updates, contradictions and forgetting.
We also sponsored the Stripe x Briefcase fintech hackathon in London, and there is more from San Francisco and Berlin further down. Next up: the Grok Bot hackathon in London on 22 October, where we are a technology partner.
An agent that matches on meaning alone cannot tell this week’s fact from last year’s. cognee 1.6.3 extracts dates when it ingests, sends time-scoped questions to temporal search and ranks results by how tightly they fit the window, boundary dates included. Check out our docs, where we run you through a sample of time-scoped questions - before, after, or between two points in time - and show you how to implement it.
Memory for coding agents. Docstrings and execution flows now go into the code graph as documents of their own, linked to the code they describe, with search and navigation built for coding agents. Instead of re-reading files every session, coding agents now have a lasting, structured memory.
Embedding costs you can see. Embedding tokens are now counted separately from LLM tokens, across every embedding engine.
Runs without a key. Copy.env.template as it is and cognee runs fully local, with local models for extraction and embeddings.
Google Calendar as a source. Self-hosted cognee now syncs calendar events into memory over a read-only OAuth connection, a week after Notion joined the sources.
cognee in Codex. The cognee plugin for Codex is live in OpenAI’s plugin directory. Codex gets relevant context on every prompt, and each session goes into your graph when it ends.
Security, written down. We published a vulnerability disclosure policy and now check for new CVEs every day. The release also brings a batch of retrieval-accuracy and reliability fixes.
Our partners at n8n hosted their first-ever In the Loop conference, and we were there with our own booth. Plenty of great talks, workshops and discussions around how to build end-to-end automations for processes in customer success, accounting, marketing, or sales. If you are an avid n8n user, check out our fully hosted, verified node on their platform. It allows you to ingest from 200+ sources into the cognee cloud.
During SF Tech Week we ran Build a Company Brain with Scalekit and Respan: an afternoon of agents that connect workplace tools, remember context and check who is asking before they answer.
Want your company to host a meetup, workshop or hackathon with us? 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.