Skip to content

feat: prioritized experience replay — revise highest-surprise memories first #79

@salishforge

Description

@salishforge

Path 1: Active Knowledge Management

Track prediction error per memory — when a retrieved memory leads to negative feedback, record the 'surprise' (expected positive, got negative). During sleep cycles, prioritize revising high-surprise memories over random low-confidence ones.

Based on Prioritized Experience Replay (Schaul et al., 2015). Closes the loop between retrieval quality and memory quality.

Design

  • Add surprise_score to warm_tier (updated on negative feedback after positive history)
  • Sleep cycle Phase 3: sort revision candidates by surprise_score DESC, not just importance
  • Track cumulative surprise per agent in health metrics

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requestfeatureNew functionality

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions