Sabitlenmiş Tweet
V1 is out at 93.8%. We've made a lot of ground since April, but we have a very specific benchmark in mind for Q3. Follow to see when we hit it.
English
CortexDB.ai by creator of Cassandra
54 posts
CortexDB.ai by creator of Cassandra
@cortexdbai
AI loses context. CortexDB fixes that. Durable memory for agents, copilots & assistants. SDKs, APIs & MCP-ready. By @pmalik
Katılım Mart 2026
6 Takip Edilen24 Takipçiler
8/Memory is the substrate on which an agent's continuity of self is built.
If it can't distinguish observation from inference, can't preserve evidence, and can't reason about time — the agent on top will be confidently wrong in ways its operators can't diagnose.
Five layers isn't a feature. It's the minimum.
English
7/What the structure unlocks:
→ Evidence traceability: walk a Belief down through Facts → Episodes → raw Events
→ Safe schema evolution: rebuild derived layers under a better extraction model, from the Events you already have
→ Genuine temporal reasoning: "What did we believe on day X, as of day Y?"
English
1/Most AI memory systems flatten everything into one layer.
When the agent tells you what it believes, you can't ask why. The chain of reasoning has been compressed away.
Here's what memory looks like when you don't do that
CortexDB.ai by creator of Cassandra@cortexdbai
English
@swyx the asterisk is where the money is.
English
harness, scaffold, context engineering, agent... do you actually know what they mean?
we wrote an AI agent glossary and tried to make sense of it all with simple definitions and real examples
↓ go read it ↓
English
The long-context discourse is cope. Recall across sessions, tools, and time is a different problem from "fit more into one prompt."
English
if you're evaluating memory layers, ask three questions: did it run through the production write path, was there an in-memory shortcut, what's the repro cost. here's ours: yes, no, $49.69. 93.8% on LongMemEval-S.
English
Agents don't forget gracefully. They forget expensively. Every re-litigated decision is a user deciding you're not worth the second turn.
English
Building long-term memory for AI agents requires knowing exactly how to handle state at scale. From @cassandra to @cortexdbai , with @pmalik on board, we've been here before.
English
We've been building CortexDB in public for 8 weeks. 498 commits. Five-layer memory model. 93.8% on LongMemEval-S, SOTA
We post every milestone, benchmark, and design decision here.
Follow @cortexdbai for the build log →
English
The piece worth defending in that pile is memory. Most of the stack is replaceable (new LLM, new harness, new framework) — the memory layer is the only part that accumulates value over time because the data it holds is yours, not the model's. Treat it as the durable layer and the rest becomes interchangeable.
English
Same thing applies in AI
> use this LLM
> wait it’s outdated, use this new one
> now you need long term memory setup ofc, your AI needs to remember things
> ok now you need this agentic harness, it’ll supercharge your LLM’s capabilities
> but you’re using too many tokens now, try this token reduction skill
> you’re only running one agent? What a scrub, you need to parallelize. Use this multi agent orchestration tool.
And on and on and on. To infinity.
And guess what .. for most people, the most effective path would have been to stop at step 1
Just use this LLM.
GIF
Aakash Gupta@aakashgupta
Every certified personal trainer alive would look at this program and call it garbage. No progressive overload, no rest days, no pulling movements, no periodization, no protein timing. Three exercises and running. That's it. For 1,095 straight days. Tasuke got more shredded than 95% of people paying $200/month for optimized coaching. Run the bull case for modern programming. Periodized splits with progressive overload produce faster hypertrophy in controlled studies. Undulating rep schemes prevent plateaus. Pull/push balance prevents injury. Rest days allow supercompensation. The science is real and the results are measurable. A good coach will get you further in 12 weeks than Saitama's routine will in 12 weeks. No question. Sounds like a win for complexity until you realize what Tasuke actually traded it for. He traded optimization for the one variable that beats all of them: a program so simple he never had to think about whether to do it. 100 push-ups. 100 sit-ups. 100 squats. 10km. Go. The $30 billion fitness industry sells periodization, app subscriptions, macro calculators, and recovery protocols because those are renewable revenue. Adherence to three exercises for 1,095 days generates zero recurring fees. You can't monetize "just keep showing up." 328,800 total reps. 10,000 miles. Built on a routine a fictional bald superhero made up as a joke. The program was never optimal. The program was never supposed to be optimal. The program was supposed to be impossible to talk yourself out of on a Tuesday morning. And that turns out to be the only fitness variable that compounds.
English