Research
The useful comparison is an agent with relevant prior experience against one without it. We want to measure the difference in outcomes, cost and repeated errors, while accounting for extra context, compute and advice.
Current questions
A previous failure could justify checking an assumption, searching another source or asking a person. How should the system choose that response, allocate its attention and compute, and make the resulting decision inspectable?
Facts, procedures and failed attempts have different useful lives. What representation, scope and expiry should each have? The record should preserve where a claim came from and what would invalidate it.
A rule that helped yesterday can mislead today. How should a system detect changed conditions, revise inherited state or reset? What knowledge should remain local because of privacy or context?
Compare against a fresh start, additional context or compute, and advice available before the experience. Measure outcomes, repeated errors, cost and escalation as conditions change. A useful experiment must distinguish help, no effect and harm.
Starting point
Agent Memory Contracts exposes source evidence, candidate claims, promotion decisions and trusted records through Python contracts, JSON schemas and validation. Its code and reference runtime are public.
It addresses the integrity of the memory boundary. Whether an agent uses that memory to make better decisions remains an experimental question.
This work draws on continual learning, external memory and adaptive control. The broader argument is in Intelligence should accumulate.