one Viktor today. at 320 million people it becomes thousands, then millions of agent instances running in parallel, each one ephemeral. put the whole swarm on one shared memory and we get network effects: every instance writes what it learns and every other instance can use it, so the more agents we run the faster the whole thing improves. wrote an essay here -xitter.cf/cortexdbai/status/2053…

RAG retrieves history. Memory maintains continuity.

cortexDB stores memory as an append-only event log, so a wrong answer comes with a source, not a shrug.

coding agents forget everything between sessions. give them one shared memory instead. local and inspectable, built for agent fleets. cortexdb.ai/download

Us

The headline finding here is quietly brutal. The gap between the two models was smaller than the gap between either model and the job. The job is experience, and experience is a memory problem. A senior analyst is a junior one plus years of remembered mistakes. This model is brilliant and permanently on its first day. It cannot remember the beta it flagged last quarter or the thesis it already watched die. The silent zero-volatility failure is the proof. A smarter model does not catch it. A model that remembers being corrected last time does. Reasoning is the model's. Knowledge is retrieval's. Experience is the layer nobody has shipped, and it is the one that turns a tool into an employee. Wrote about it here - xitter.cf/cortexdbai/status/2054…

Yes, exactly. Retrieval stops being enough when we move from answering to acting. Relationships become the substrate. And those relationships live in time. Not just what connects to what, but when it was true and how a belief formed. Give the graph that, and an agent can do more than traverse the world. It can reconstruct how the world got this way. That's the part of memory we keep coming back to. Wrote about it here: - xitter.cf/cortexdbai/status/2063…

captcha

"AI teammates, not chatbots" and "safe for enterprise" are one requirement. a teammate is an agent whose memory persists and beliefs trace to the events behind them. legibility makes them deployable xitter.cf/cortexdbai/status/2063…

6/6 trust used to be asserted: a benchmark, a brand, a promise. traceable memory makes it queryable. "show me the events under this conclusion." cortexdb is that bet: every input stored as an immutable event, never rewritten, everything else derived on top.

5/6 this is becoming procurement, not a feature. 88% of orgs hit an ai-agent incident last year, only 21% could see what their agents were doing. eu ai act article 12 makes event logging mandatory for high-risk systems in aug 2026.

4/6 the honest limit: reading the memory store is not reading the model. we can see what the agent knew. we can't fully see why it acted. a wrong belief is auditable. a wrong action built on correct beliefs is not, yet.

3/6 so the first cognitive system whose reasoning can actually be read won't be human. it's readable precisely because it lacks the interpreter that fills our gaps silently. ask why a belief is held and either the chain is there or it isn't. the absence is visible.

2/6 every belief humans hold comes with a story assembled after the fact. the chain that actually produced it was never stored anywhere we could read. agent memory is the first tech that stores it.

TLDR: 1/6 ask anyone why they did something and you get a fluent, confident answer. it can be entirely made up. the split-brain work showed it at the wiring level: the brain invents a reason, reports it as the cause, and never knows it's inventing.

Selective Amnesia cortexdb.ai/

A man of culture, I see

agents ARE rebuilding software platforms for other agents and cc'ing us on the launch email out of politeness.