We have 85 PRs merged last week and some big announcements: - cognee 1.6.0 release now runs without an LLM API key: local GLiNER extraction + fastembed embeddings (demo mode) 🔥 - Cloud: GitHub connector landed, connect your whole org and load all your projects in minutes +  live-streamed search answers - improve() became an explicit orchestrator, and recall routing got stricter — unresolvable dataset names now fail loudly instead of being silently dropped - Typed edges for custom graph models + Presort: folder pre-organization and a JSON graph-model DSL - Security hardening: random defaults replace the hardcoded token secret and default-user password, external ref schemas rejected, MCP transport origin guards - cognee-mcp: PyPI Trusted Publishing, X-Api-Key auth for self-hosted mode, recall results now summarize memory hits

Our Research Lead did not write a line of code in months We write plans. These plans become POCs which become new plans. Once we are happy, we harden the POCs, add tests and release them to production At every stage there is an expert. Data Science, Product, Engineering

Reminder to join our webinar on how to build a company brain, Tuesday 22 September. 45 min, online and free. Which data sources to connect on day one, and what your agents can ask the graph once it's built. 20:00 CEST / 11:00 PT

Memory related posts keep going viral because people can actually feel the problem! As a leader in the space building @cognee_, we wrote a detailed guide which gets deep into what agent memory actually is. Read more here:

hot take: To promote a startup you will need to win a battle of bots in the future on social media. Distribution channels where bots will act like bots in financial markets, spread the rumours, and then sell the news, and hack things in the whole financial markets.

Great to see that the community finds our blogs useful! This guide is a practical map of the benchmarks used to evaluate AI memory today: what each one tests, how it's constructed, and where to find the paper and data. We also discuss the gaps in existing benchmarks and how to evaluate memory for your own system. Go check it out and let us know what you think!

Your agents don't know what your company knows. Tuesday 22 September I'm doing 45 minutes on how to build a company brain, live on @cognee_ . Slack, GitHub, Linear and your drive into one graph your agents can query. 20:00 CEST / 11am PT, on Zoom:

What we shipped at @cognee_ last week → Cloud pricing dropped to $1 per 1M tokens → LLM-free graph extraction landed (GLiNER) - build knowledge graphs without burning tokens → Coding-agent plugins (Claude Code, Codex) got faster recall: concurrent scope dispatch + per-agent identities with scoped permissions → Activity log now shows per-operation cost, and memory coverage scores only what was actually measured → Docs overhaul: fixed API playground, Rust & TypeScript SDK tabs, new pipeline demos 151 issues closed across the team. Shipping memory that works.

Great to see @sama and @DarioAmodei to agree on something! Especially after the recent Anthropic Threat Report, which btw was one of the most insane articles I have read recently. If these were reported incidents, you can only imagine how many other exist that we don't know about... Quick TLDR: - Alibaba extracted 151M+ Claude exchanges to help train Qwen. - Moonshot and DeepSeek routed some users to Claude without telling them. - A weapons group used multiple Claude instances as an engineering team, test-fired a guided rocket, failed, and then returned to Claude to understand why. Let's be honest... we don’t hear about such incidents until a someone publishes a report, or someone inside decides to speak publicly... Quite concerning imho, and to be honest with you, a good solution does not exist as of today.

Shipped last week at @cognee_ : - v1.5.4 released - Streaming recall over SSE (opt-in) - cognee-cli demo — full graph demo, zero API keys - add() takes raw text + repo URLs directly - Chunk-level incremental loading - Strict ontology grounding (opt-in) - Signed release provenance + hardened CI - Explicit plugin memory access management — you now manage what a plugin's agent can read/write instead of implicit access. - opencode got a native integration — lifecycle hooks, recall, and durable capture. - Security fixes: removed watcher credential exposure and the prompt-identity delay; bounded recall and protected automatic capture from runaway behavior; bounded cold-start retries per conversation in openclaw

Last two weeks I shipped myself 46 tickets, most closed and merged A few key things: - Memory now adapts to each user: preference nodes with weighted "prefers" edges personalize retrieval per user in the same graph, different answers depending on who's asking. -> great help from Lazar - recall() on a not-yet-built graph now returns instantly with a "memory warming up" marker instead of running a full graph search + LLM call. We found 511k wasted searches from empty-graph queries - New cognee-cli doctor: misconfiguration now fails loudly at first use with preflight checks, instead of surfacing as a cryptic error three steps later. - Every fact in your knowledge graph can now carry an audit-grade provenance trail (W3C PROV-O)