Tiered Memory
Memory split by access cost: a small always-loaded hot tier (digest, index), a warm tier loaded on relevance (fact bodies), and a cold tier queried explicitly (archives, logs).
When to use
Keeping the per-session context tax flat while total memory grows without bound.
Specification
| Memory type | semantic |
|---|---|
| Storage | layered: file index + fact files + archival store |
| Retrieval | hot always, warm on relevance, cold on explicit query |
| Update method | writes land warm; promotion and demotion by usage and age |
| Decay / staleness handling | Demotion is graceful decay: unused facts sink toward cold instead of being deleted, preserving retrievability without paying context cost. |
| Risk level / maturity | low / emerging |
Privacy and safety
Per-tier access rules; cold archives especially accumulate everything and deserve their own controls.
Failure modes
- Promotion logic that never demotes, inflating the hot tier
- Cold tier becoming a write-only graveyard
- Tier boundaries invisible to the agent, causing missed recalls
Assessment
What indexed fact files grow into when they succeed. The design rule: the hot tier is a budget, not a category, and everything fights for its place there.
Evaluation method: Hot-tier size trend (must stay flat) and cold-tier hit rate on explicit queries.