Decisions
When should a previous outcome change the next search, test or action?
Building systems for cumulative intelligence
Agents should improve with experience.
We build memory, provenance and evaluation infrastructure for agents working with changing information.
Explore the workThe problem
Memory is not enough.
A stored answer can be outdated. A successful procedure can fail when the task changes. Keeping more history does not tell an agent what to trust or do next.
The challenge is to connect retained experience to decisions, and to notice when the past has become a liability.
Public technical work
Agent Memory Contracts
An agent’s interpretation should not silently become a fact. Agent Memory Contracts is a public Python library for making that boundary explicit.
Source evidence stays linked to a candidate claim. Accepting the claim requires a recorded decision, and the validator rejects a trusted record without the required authorization. The worked reducer example shows acceptance, rejection and an invalid promotion.
What we are investigating
01 / 04
Useful state
Which facts, procedures and failed attempts deserve to persist, and with what source and scope?
Revision
How should a system detect stale knowledge and decide what to revise or reset?
Evaluation
Does retained experience improve results when context, compute and prior advice are accounted for?
Repeated work
Bring a problem that keeps coming back.
Does an agent reconstruct the same context every week, repeat a known mistake or fail when a fact changes? We want to work on recurring workflows where the result can be checked and a person can intervene.
Tell us about your workflowThe broader thesis
Intelligence should accumulate.
Learning costs time, attention and failed attempts. The essay asks what should survive that investment, how it can pass between systems and when it should lose authority.
Read the essay