Context the whole company can use.
Can one brain improve the whole workforce?
Remembering is not retrieving
Most systems described as agent memory are retrieval systems: text goes in, similar text comes back. That is useful, and it is not the same problem. An agent that worked with you last month should know what was decided, what changed since, and what is no longer true — none of which is answered by finding a similar paragraph.
We model memory as claims with two clocks: when something was true, and when we came to believe it. That lets a store answer what it believed last Tuesday, and lets a correction supersede a fact without deleting the record that it was once held.
A negative result we keep repeating
The intuitive design is to have a model read every conversation as it happens and write structured facts out of it. We measured that and it did not earn its cost. The failure is structural rather than a tuning problem: extraction decides what to keep before anyone knows what will be asked, and what it discards is unrecoverable.
The lever that did work sits on the read side. Deciding how to interrogate a store, at the moment of the question, was worth more than anything we did at write time. Both matter; the negative result is the more useful of the two, and neither lands on /lab/papers until the research gate clears.
One store or many
If every agent keeps its own memory, a company learns nothing — each agent relearns what the last one already knew. If they share one, the questions become who may read what, who may write, and what happens when two agents disagree about the same fact.
We are building toward shared, with the disagreement handled explicitly rather than by last-write-wins. Whether that holds up under many concurrent writers is genuinely open.
- Status
- Research
- Focus
- Cost and usefulness (research)
- Surface
- Early / evolving
- Papers
- Findings land on /lab/papers when gated