Make memory
accountable.
Our core question: can a condition-aware, fail-closed memory-governance architecture preserve useful operational experience while suppressing stale, contradictory, weakly supported, mis-scoped, or unsafe memories as conditions change?
What we study
We develop general-purpose methods for preserving, evaluating, governing, and reusing verified operational knowledge in AI-assisted systems. This is a decision-control problem—not another language model, document repository, or application workflow.
- Versioned operational-memory representations
- Evidence and provenance binding
- Human verification and review states
- Contradiction and stale-condition detection
- Uncertainty-aware retrieval and synthetic adversarial evaluation
Stricter controls need not destroy usefulness.
We test the utility-versus-suppression frontier instead of assuming either permissiveness or silence is optimal.
Provider-neutral by design.
Model output is an experimental recommendation object, never the system of record.
Fail cases are first-class evidence.
Stale procedures, conflict, incomplete evidence, prompt injection, and near-duplicate events are deliberately introduced.