Most memory products conflate three different questions into a single “delete” or “rank lower” behaviour. That is why teams discover too late that a correction was lost, a stale preference still wins at retrieval time, or an erasure request left records in a backup table nobody queries.
Spectron separates them on purpose:
| Question | Mechanism | Example |
|---|---|---|
| “This replaced that.” | Supersession | You preferred formal tone in March; you prefer casual now. The old instruction stays in history with a bounded validity interval; the new one is current. |
| “This stopped being relevant.” | Decay with reinforcement | A sprint goal from three weeks ago fades unless retrieval keeps using it. |
| “This should not exist for the agent.” | Forget (soft by default) | “Forget everything about my old job.” Matched beliefs expire from current retrieval; see Forgetting memories. |
Conflating these into one overwrite or one ranking pool is a common reason agent memory fails in production: you either lose the audit trail or you let two “current” beliefs fight at query time.
Supersession — replacement with history
When an updated preference or corrected fact arrives from the same kind of source (for example two statements from the same user over time), Spectron treats it as you changing your mind, not a contradiction to flag.
The prior belief is closed in time, not erased.
The new belief becomes the single current answer for “what is true now?”
You can still ask what the value was last month and when it changed.
Instructions behave the same way: “always use formal tone” superseded by “prefer casual” means the agent follows casual now, but the shift remains auditable.
Cross-provenance clashes are different — an uploaded policy versus a user assertion — and surface as uncertainty instead of a clean supersession. See Authority when pillars meet.
Decay — relevance without destruction
Not everything should live at full strength forever. Context fades quickly; knowledge fades more slowly; identity is comparatively stable. Spectron applies category-aware importance decay on a schedule.
Decay is not blind deletion: when a memory contributes to a successful retrieval, its importance can be reinforced so useful facts stay sharp while untouched noise quietly leaves the active set.
Background consolidation can also crystallise repeated observations into durable beliefs (source.kind = "consolidation") — the “between conversations” evolution described in Traces and memory evolution.
Forget — explicit removal
When something must stop influencing the agent — privacy, compliance, or a hard reset — use POST /forget or entity delete. Default behaviour expires matched records from current retrieval while keeping audit history; purge: true and scoped forget exist for erasure requirements.
Forget does not block future extraction: if a later turn mentions the same topic again, new memory can be created. Combine product policy with forget when a topic must stay gone.
Long sessions and compaction
Existing model context windows encourage compaction, namely the process of squashing a long thread into a summary. Summaries are inherently lossy, often discarding details you did not know you would need later.
Spectron’s approach is complementary: important facts from turns are promoted into durable, queryable memory (entities, attributes, instructions) while the full episodic transcript remains addressable via sessions and provenance. Consolidation and reflection run between interactions so you are not dependent on remembering to compact at the right moment for memory to survive the next session.
Who owns the decision?
The reconciler owns the bookkeeping guarantees — supersession chains, confidence floors, decay schedules, uncertainty when sources disagree. Signals come from agent actions: turns, corrections, document ingest, reflection, and traces of what was retrieved and used.
When the system genuinely cannot decide (cross-provenance conflict, confidence below the calibration floor), it records uncertainty for the agent or a higher-level policy to resolve rather than silently picking a winner.