Memory research

Useful memories should earn their recall budget.

Static importance is a good starting guess, but real utility shows up later. Memory Worth tracks whether recalled memories actually helped, then uses that signal to spend context more wisely.

See all memory features How memory works Issue #560
This is one research track, not the whole Remnic memory system. Remnic also includes extraction, recall injection, hybrid search, entity tracking, lifecycle hygiene, trust zones, procedural memory, local LLM routing, versioning, retention tiers, disclosure controls, project scoping, importers, benchmarks, recall X-ray, and more. These pages explain focused memory tracks inside that broader system.
What changes

The product idea in plain terms.

Keep static importance

Write-time importance remains the prior. Memory Worth adds runtime evidence instead of pretending the first score was perfect.

Track help and harm

Recall audit signals increment success, failure, or unknown counters for each injected memory.

Filter softly

Low-worth memories are deprioritized rather than deleted, preserving auditability and preventing premature forgetting.

Already in Remnic
  • Importance-gated writes keep trivial chatter out.
  • Recall audit records which memories were injected.
  • Lifecycle and retention policy work now separates hotter and colder memory tiers.
Landed in this track
  • PR #573 added Memory Worth counters to frontmatter.
  • PR #616 added computeMemoryWorth, PR #620 wired outcome signals, and PR #626 added the recall filter.
  • PR #628 added benchmarks and flipped the default on.
References
Related Remnic pages