Skip to main content

Forgetting Strategies

Memorose is designed to forget on purpose. The goal is not maximum retention. The goal is durable, useful retention.

L0: Entropy Filtering​

Low-signal events can be filtered before or during consolidation using the entropy threshold configured under [consolidation].

This keeps greetings, filler, and repeated low-value chatter from turning into long-term state.

L1 And L2: Decay And Pruning​

Long-lived memory should lose weight if it stops being useful. The [forgetting] section controls:

  • decay half-life
  • minimum importance
  • pruning interval

This is what prevents a long-running agent from turning into an unbounded junk drawer.

Graph And Community Compaction​

Graph auto-linking and community detection let the system replace repeated local evidence with higher-level structure when that structure is more useful than replaying every source fragment.

L3 Sedimentation​

When a task completes, its outcome can sediment back into the event stream as a new fact about what happened. This lets execution outcomes become memory instead of staying trapped inside task state.

Operational Principle​

Memorose treats forgetting as part of the architecture, not as a cleanup script that runs after the design is already wrong.