Mark consequential transitions.
Connect decisions to state changes so later retrieval can answer “what happened because of this?”
Case study · Adaptive agents
Memory that changes the next move.
A useful memory is not an archive of everything. It is a working representation that preserves consequential evidence, exposes uncertainty, and helps an agent choose differently when a familiar situation returns.
- Status
- Active research direction
- Central question
- What should an agent remember when context is long but decisions are immediate?
- Evidence standard
- Behavioral change, controlled ablations, calibration, and measured resource cost.
01 · Question
Which past details deserve a future vote?
Long interactions create more observations than a system can keep active. Retaining every token is expensive and can bury the decisive event; compressing too aggressively can erase the exception that matters. The useful middle ground is memory organized around decisions, causes, and unresolved questions.
We study how a memory can distinguish stable rules from one-off events, connect an outcome to the action that caused it, and surface contradictions instead of smoothing them away.
02 · Approach
Store consequences, not just chronology.
Keep summaries inspectable.
Retain provenance and uncertainty around a compact memory instead of presenting a summary as settled fact.
Ask from the current decision.
Retrieve evidence relevant to the next action, not merely the text most similar to the latest prompt.
Make contradiction productive.
When new evidence conflicts with memory, preserve the disagreement long enough to revise the model.
03 · Evidence
The test is whether behavior improves.
This page describes a research program, not a released benchmark result. Evidence for an adaptive-memory system should show that the memory changes later decisions in the intended direction—not simply that stored text can be retrieved.
Counterfactual comparisons
Compare the same task with memory removed, truncated, or replaced to isolate what the memory contributed.
Confidence under contradiction
Measure whether the system lowers confidence and seeks evidence when remembered rules no longer fit.
Cost per useful decision
Report storage, retrieval, latency, and context overhead alongside task outcomes.
Traceable influence
Make it possible to inspect which memory informed a decision and how it was formed.
04 · Limitations
Remembering can amplify the wrong lesson.
A memory system can preserve false assumptions, overfit to unusual episodes, expose sensitive context, or create a convincing but inaccurate narrative of cause. Performance on long-context retrieval alone does not establish durable learning or safe behavior.
Current boundary. We describe evaluation criteria and design questions here. We do not claim a generally reliable memory system, production availability, or validated transfer beyond the environments in which a future result is measured.