Skip to main content
Dream consolidation is an opt-in runtime mechanism that recombines selected, temporally distant memories. It can write journal material, low-confidence beliefs, and tagged GEN fragments. It is separate from routine memory consolidation and from context-window compaction. Its outputs are generated associations, not newly observed facts.

​
Default state

Dreams are disabled in the checked-in harness configuration:
When enabled, the shipped defaults require at least an hour of silence, at least four hours between cycles, and permit at most two cycles per day within the configured local-hour window. Deployment timezone therefore affects eligibility.

​
Data flow

The dream engine selects distant memory pairs, requests creative recombination, and routes the result to the enabled destinations. Configuration controls journal writes, belief writes and confidence, GEN injection, and output limits. Keep the distinction between the original memories and the generated connection. The default belief confidence for dream-derived material is deliberately low (0.35), but a numeric confidence value is not an external truth assessment.

​
Operational impact

Dream generation uses inference during idle periods. Enabling it changes the worker’s background usage even if no one is chatting. Use the shared inference pause to stop it along with other model consumers. Inspect resulting journal entries, belief records, and tagged fragments when evaluating the mechanism. Compare against a run with dreams disabled before attributing a behavior change to this subsystem. Source: ngram/identity/dream.py.