Retire, Redact, Supersede, And Relax Memory
Type: kb/types/note.md · Status: current · Tags: agent-memory, context-engineering
Learning is incomplete without forgetting and revision. Raw traces can contain secrets or obsolete assumptions. Observations can be duplicates, wrong, low-value, or superseded. Cues can grow stale. Policies can become too broad. Tests can fossilize temporary workarounds.
Append-only capture is useful for provenance. Indexes, extracted observations, and activated policy surfaces still need redaction, decay, supersession, retirement, and relaxation. This is context efficiency as lifecycle work: every stale artifact competes for attention, search rank, review time, or behavioral authority.
Methods
- Retention classes and redaction status on traces before model extraction.
- Candidate, accepted, superseded, rejected, and retired states on extracted observations.
- Periodic triage for observation inboxes: promote, fold into an existing artifact, keep, reject, or delete.
- Duplicate clustering and source consolidation for repeated observations.
- Recency decay tempered by consequence and recurrence, so old high-impact corrections do not disappear merely because they are old.
- Retirement tests for cues that fire often but do not change behavior or produce too many false positives.
- Relaxation paths from rigid enforcement back to prose guidance when a codified rule proves brittle.
Temporal Memory
Some domains also need queryable temporal memory, not only lifecycle labels. Supersession answers "what is current?"; temporal memory answers "what was true then?" Graphiti's bi-temporal edge model, summarized in the comparative review, is the clearest reviewed case: validity and invalidation timestamps support point-in-time queries and temporal contradiction handling. Flat files can approximate this with provenance, dated decisions, and git history, but systems that need temporal recall should treat time as a query dimension.
Related Systems
Reviewed systems show both partial fulfillment and a common failure mode. cass-memory has candidate/established/proven/deprecated states, harmful-feedback weighting, and decay. REM defines active, contradicted_by, and superseded columns but does not wire them into an actual update path. Lifecycle metadata becomes memory management only when some process reads and acts on it.
Evaluation Questions
- Does lifecycle metadata drive actual processes?
- Can the system redact sensitive traces before extraction or reuse?
- Are obsolete cues, policies, and generated views retired or relaxed?
- Can the system answer temporal questions when current-state memory is insufficient?
Relevant Notes:
- Context efficiency is the central design concern in agent systems - explains why stale memory has a real cost
- Designing a Memory System for LLM-Based Agents - derives lifecycle from context pressure and source authority