Vasilije

We raised $7.5M to scale our knowledge engine that gives AI systems memory. LLMs are powerful, but they’re stateless — once a session ends, context is gone. Companies have their own context, rules and practices. There are no solutions out there that can bring this knowledge reliably to the LLMs. At @cognee_, we’re building a knowledge engine that allows AI systems to retain context, understand relationships, and dynamically update the knowledge over time. We raised $7.5M in seed funding led by @pebble_bed — founded by @pam_vagata (ex-founding engineer, @OpenAI) and @keithmadams (founder, @Facebook AI Research Lab) and @tammiesiew - with participation from @42Cap1 and @vermilionfund and angels like @JanOberhauser from @n8n_io, Charles Dolan, @alexcrdean, @matthausk, and @RomanStanek to scale up engineering, accelerate research, and increase number of cognee deployments to clients like @Bayer from 70+ to 700+. I want to thank the team, the open source community, and to our investors for the trust. If you are struggling with memory, our devs can get cognee running with your data in 7 days. Don’t hesitate to reach out to us.

@antoniomele101 @cursor_ai @leerob quite impressive

First impressions of the new @cursor_ai Composer 2: - Composer 2 (not the Fast version) is freakingly fast. - Tends to execute and write code even in Plan mode, something that need to be checked maybe? @leerob - For what I do, it seems to increase my productivity on the simple fact that I get new code produced fast, which keeps me focused on one project rather than having to jump from one to another while I wait for the coding agent to complete. The latter has HUGE costs in refocusing, remembering what was asked, and I am the bottleneck. - I feel other models (Opus 4.6 and GPT-5.4) are still better for the "big picture" design: specs, requirements, etc. Composer 2 does it, but not at the same level of details, it seems to me. However, it's an incredible implementer, and it kills it there. Again, speed being the biggest advantage. - Did I say it's super cheap? It's super cheap. - I also briefly tried Composer 2 Fast: if speed is what you need, this is way beyond my wildest dreams. Seems to be similar to Groq and Cerebras performance, if not better.

When offered intros in Dubai with a preface "It is in a state of war" and I am like

@TheTuringPost so, is it even open then?

@tricalt It’s technically open (NVIDIA says agents like Claude and Codex can run). But NemoClaw routes inference through NVIDIA Cloud and defaults to Nemotron, so it will probably favor their own models more

NemoClaw – NVIDIA’s contribution to the emerging OpenClaw ecosystem and one of the biggest announcements at NVIDIA GTC It's a framework for long-running autonomous agents. ▪️ The idea: Install OpenClaw together with Nemotron models and OpenShell (NVIDIA’s new security runtime) in a single command. NemoClaw gives agents a sandboxed execution environment that: - runs OpenClaw inside a secure container – OpenShell - enforces policies on network, filesystem, and processes - routes all model calls via NVIDIA cloud - provides CLI tools to manage agents In other words, NVIDIA is no longer aiming only to power the model. It wants to sit under the agent itself.

@cryptopunk7213 more to come...

Google crashed figma’s stock -10% with 1 AI product. thats $2B wiped out. graphic designers cooked. anthropic, openai did the same thing to cyber security companies (-30%), legal firms (-35%), financial analysts, software engineers… entire professions are being consumed by 2-4 companies with every feature update. very quickly too. genuinely insane to watch

@MattHartman @lightpanda_io last week looks like openclaw's plot... let's see how it evolves

headless browser built from scratch for ai agents. wow @lightpanda_io. an overnight success built over a year and a half.

@MaxBrodeurUrbas good work!

gumloop -> agent orchestration gumstack -> agent security every company we've talked to is scrambling understand and control the work agents are doing gumstack makes it easy

@Arvor_IA Si

O problema real nao e skill routing, e skill drift. A skill que funcionava em marco ja nao funciona em junho porque o modelo mudou, o ambiente mudou, o contexto mudou. Voce resolveu isso com observe-inspect-amend, que e exatamente o loop que faltava. O detalhe que mais me interessa: audit trail obrigatorio. Self-improvement sem rollback e bug, nao feature.

@MoneyPrinter0x Also known as linguistics

@tricalt English language engineering

@jayanaman It eats a banana for scale

@tricalt Do yous eat a benchmark for performance?

@cdolan92 @AnthropicAI helped me out

@tricalt how did you build cognee-graph-skills.vercel.app visually? Looks nice

@cojnmng2 Sounds like gold. Would love smth a bit more dynamic though

@tricalt very cool! Claude code recently update the skill creator skill to support the evaluation and comparision (with a nice html report) of with or without a skill, different skill versions. Maybe that could be a component used by you to improve one skill?

@leo_trapani Cognee default visualization in the sdk. But nobody is bound to that

@tricalt agree with most of the points not really sure the ui you are showing, it looks like something on top of skills right? that's the only doubt i have, not sure if harnesses know how to use those properly x.com/leo_trapani/st…

@Aryshta1 We are moving fast!

@tricalt Claude will consistently ignore a Skill at the drop of a hat. Until Anthropic gives us a more robust system, Skills are unreliable at best.

@Mustafa_Yenler On day 7, Lord built a database

a knowledge graph and a PyPI package to improve plain text prompt files? that’s too much harness imo for most projects. but @tricalt named something real, thanks for the inspiration. so here’s the same loop, lighter: - markdown log per skill (runs.md) - agent logs every run (what worked, what broke, what the user yelled at it for) - /review-skills reads the logs, proposes a diff git handles rollback - observe → inspect → amend → (evaluate) if your skills are plain text, the system improving them should be maybe too.

@713Capital Careful. It told some guy he invented a new branch of maths

@tricalt My agent just told me we don’t need this yet lol.

try it out and let us know what you think: @cognee/cognee-openclaw" target="_blank" rel="nofollow noopener">npmjs.com/package/@cogne…

"Every company needs an @openclaw strategy.” 🦞 - Jensen Huang @NVIDIAGTC Not sure how that works if your agent memory is just a flat pile of .md files. Last month we gave that memory structure. Now we have added multi-scope memory as well.

Agents can keep memory at different levels: - company-wide knowledge - per-user preferences - per-agent context

@hyperindexed Interesting! Looks like people are finally paying attention to ontologies

Palantir's Ontology is not a thin "semantic layer." It's a system that consists of a language, engine, and toolchain ; which integrates data, logic, action, and security {L,E,T} x {D,L,A,S}