The standard memory story for AI agents goes like this: the model has a context window, the window fills up, and everything that scrolls out of it is gone. Vendors patch this with retrieval stuff documents in a vector store, pull back whatever looks similar to the current question. It's better than nothing. It is also not memory. It's a filing cabinet bolted to a patient with amnesia. The patient can look things up, but he doesn't remember deciding to file them, doesn't know what's in the cabinet, and can't tell you what happened last spring unless you guess the right search terms on his behalf.
We built Cortex by asking a different question: what does memory actually do, in the systems where it works?
What human memory gets right
Three things, mainly. Recent events are sharp you can replay this morning's conversation nearly verbatim. Older events compress into narrative you don't recall each standup from last quarter, you recall "that was the quarter we rewrote the billing system," plus a few durable specifics. And detail is recoverable on demand: given a reason to dig, you can descend from the narrative back toward the moment. "When did we decide that? Sometime in spring. Around the offsite. No the week after, in Dana's review." Each step narrows the range until the detail surfaces.
Compression isn't a flaw in this system. It's the feature that makes decades of experience fit in a head and stay useful. The flaw would be compression without a way back down.
The temporal ladder
Cortex organizes an agent's history as a temporal ladder. The bottom rung is the raw record: every event, journaled as it happened. Above it, each rung holds the same history at a coarser grain days summarized from events, weeks from days, months from weeks, on up to a standing narrative of the whole relationship. Recent rungs are dense; distant ones are compressed. Nothing is deleted when it's summarized. The summary is a rung, not a replacement.
What an agent loads by default mirrors what a person carries by default: the top of the ladder, plus sharp recent detail. That's cheap, and for most work it's enough. The rest is one descent away.
Recursive recall
Recall in Cortex works the way the "Dana's review" example works. A query starts at a high rung "sometime this spring, something about the vendor contract" and the matching narrative points down the ladder to the weeks it summarizes. Those point to days, days to events, until the agent is holding the exact moment: the message, the decision, the reasoning attached. We call it recursive recall because each level's answer is the next level's question. The agent doesn't need the right keywords for a five-month-old event. It needs what you'd need: roughly when, and roughly what it was about.
A durable memory for Neo
Neocortex gives Neo temporal continuity across conversations and restarts. Decisions, preferences, evidence, goals, and open work live in durable state rather than depending on one context window. Workmates and Workforce use different state models and should not be described as sharing one universal agent brain.
Integrity
Memory this durable has to be trustworthy, and we hold Cortex to three rules. The journal is tamper-evident: entries are cryptographically anchored, so history can be added to but never quietly rewritten an agent's account of the past is checkable, including by the agent. Secrets are never stored: keys, tokens, and credentials don't enter memory, period; custody is Paxport's job, and memory remembering a password is a leak with a long shelf life. And a memory is never fabricated: if recall comes up empty, the answer is "I don't have that," not a plausible reconstruction. A memory system that fills gaps with fluent guesses is worse than no memory at all, because you can't tell which kind of answer you got.
Amnesia with extra steps was never going to carry a permanent agent. Memory that behaves like memory might.