cognee 1.6.7 is out this week, with around 180 pull requests across 20 repositories. - cognee v1.6.7 makes ingestion safer by default. Raw database connection strings and server file paths over HTTP are now refused unless you explicitly opt in, and the same switch that gates local paths now gates S3 inputs. If you ingest database sources on purpose, check the opt-in docs before upgrading. - Search got more precise. Typed and scoped queries are honoured consistently in chunk search, quoted scoped phrases skip the LLM entirelyRetrieval scoring is batched on PGVector, Turso and LanceDB, so multi-query recalls make a fraction of the round trips. - The Rust SDK landed its session-turn model: structured session persistence, a guidance block with served-context lineage, concurrent retrieval merged under a turn lock, and recall evidence with streaming answers backed by an edge-evidence ledger. - cognee-mcp 0.5.8 adds references with dataset provenance on recall, history-only forget, and ontology uploads in any RDF format. The Codex plugin now has a proper listing in the ChatGPT app. - Cognee Cloud 1.6.6 and 1.6.7 brought the new dashboard and Memory Review to production, a welcome page that fills in as your workspace is set up, and Obsidian vault import.

For the past year, we have done a lot of work with clients. We have worked with everyone from small startups to large enterprises. It is a tough job, finding the right person to talk to, convincing the budget holders, proving value, and making sure that, in their custom environment, the product works as they expect it to. When the job is done, we leave, and we don't ask for much. We ask for a quote. And sometimes you don't even get that. You just learn that the communications department is refusing to answer emails. The irony. This is to all the silent heroes of the year. There are a few. You won't be known on Twitter or LinkedIn, but Veljko Kovac and I will know.

Finally - I’m enough weeks in to using Claude Code myself in anger on multiple parallel projects to have my own take on all of this as a coder, vs my views from the day job. I am not going to wade in to the slop wars - there are many people smarter than me with better takes, shoutout @dexhorthy, shoutout @GeoffreyHuntley, shoutout @vaibcode. My take is that @Steve_Yegge’s maturity curve is real, but it’s more of a hero’s journey than Thetan Grade IV. I won’t link to the 8 stages because the stages don’t matter - everything is moving too fast. But the hero’s journey 🌋 for a software engineer with Large Coding Models is real. I can reason from first principles that I will end up at factories - but I need to surf that maturity curve myself. I want to watch the whole boxed set. I want to enjoy each season finale & villain. I want to run 6 Claude Codes in parallel, I want to buy the third screen & dedicated voice recorder, I want to run OpenClaw, I want to write my own harness, I want to run Rex from @superlogical . What does this mean more broadly? There will be some programmers who don’t want to leave the shires 🌳. They will be okay for a few years, because the Gibsonian future is unevenly distributed. There will be engineers who try and speedrun straight to the $current_end_state, the best of us can do this. For everybody else (for me) - they will want to go on the journey. They will want to have the watercooler moment and share each surprise & delight. They will want to bear witness to the biggest change in their professional lives themselves, personally. You will know they are on the journey - *their* journey - because the bags under their eyes will have bags, because you will get pretentiously worded PRs from ‘them’ at 2am. You’ll be on a Zoom and they’ll keep looking away for 1.5 secs. They will bring out an Alienware laptop in a crowded pub and forget to finish their Guinness. Let them cook.

cognee v1.6.6: 200k lines of code and 150 PRs - Conflict resolution is now a first-class workflow - Building memory got cheaper. Chunk summaries are LLM-free by default - Encrypted at rest. Together with per-user search-history deletion from the RC, this closes two of the most common questions we get from security reviews. - Open standards for the graph. You can now query any dataset with SPARQL, export an ontology in Turtle, RDF/XML, JSON-LD or N-Triples, and turn on OWL-RL reasoning so subclass, inverse and domain/range relations are honored when grounding extraction. Deletes return a receipt stating what was removed. - Recall routes smarter. Questions route across datasets by vector similarity, code and count questions go to the right engine, and routing respects whether an LLM is available. Plus everything from the release candidate: Notion, Obsidian, Asana, OneNote and Apollo connectors, faster resumable Drive sync, and user-controlled search history.

@n8n_io hosted their first conference, In the Loop, in Berlin. It was an amazing event, and a pleasure to meet everyone who joined, talk to the organizations attending and have so many great chats with you. Thanks to the @n8n_io team for having us, and to everyone who stopped by. Looking forward to future collaborations!

cognee v1.6.6 RC is out. 55 PRs merged - Five new connectors: Notion (with the page tree as graph edges), Obsidian vaults, Asana, OneNote, Apollo. Google Drive sync is faster and resumes after interruptions. - History you control: delete your own search history via the API, forget(clean_history=True) to purge search traces on their own, and deleting a dataset now takes its history with it - Faster at scale: HNSW indexes on pgvector, Ladybug bulk writes on the primary-key index, one edge lookup per batch instead of one per edge. - Traceable answers: every recall hit now carries its dataset and source data id - Plus: the global context index is now built without an LLM call, budget/quota errors surface consistently instead of silently retrying We also fixed many security issues along the way in our SOC2 preparation.