All research systems
Conflict LensExperimental

How should a memory store decide what a new fact means relative to what it already holds?

Question → hypothesis/design → architecture → evaluation → results → limitations → next questions

Motivation

Contradictory memories rarely announce themselves. They surface as confident, wrong behaviour. Making disagreement visible and inspectable is a prerequisite for trusting a long-lived agent's recall.

Hypothesis

Classifying each new fact as an add, an update that supersedes an old fact, or a duplicate stops contradictions from accumulating; precision and recall still need measurement.

Architecture

  • Dependency-free Go library extracted from MemKit
  • Token-overlap thresholds classify a new fact against stored candidates: duplicate, update (supersede the closest, keeping it as history) or add
  • An optional model-backed resolver for the ambiguous middle band

Evaluation methodology

Planned evaluation: precision and recall of the classification on MemKit's conflict workload. No scored result is reported here.

Results

Evaluation in progress

Limitations

  • Token overlap measures wording, not meaning; the resolver exists for cases the thresholds cannot settle

Current status

Released as v0.2.0; MemKit's conflict resolution is built on v0.1.0.

Roadmap

  • Scored detection benchmark
  • Reviewer workflow for bulk resolution

Next questions

  • Belief revision for agent memory
  • Cost of contradiction: linking stale recall to downstream action errors