Long-term AI memory inevitably accumulates contradictions. Users change preferences, correct themselves and sometimes speak hypothetically. A memory system therefore needs conflict resolution, not just retrieval.

Classify the source of each memory

Explicit user statements should be distinguishable from model inference, imported profile data and behavioral signals.

Use recency carefully

Newer information is often more relevant, but a recent joke should not overwrite a stable verified fact.

Explicit correction should dominate

When a user says a remembered fact is wrong, the system should supersede it and prevent the stale version from resurfacing.

Preserve provenance

Version history allows systems to explain why one memory replaced another without retaining unnecessary raw conversation.

This extends the lifecycle questions in Relationship Memory in AI Companions.

Bottom line

Reliable memory requires confidence, provenance, correction and user authority over conflicting information.

Identify the type of contradiction first

Not every conflict means one memory is false. A user may have moved from Tokyo to Osaka, changed a hobby, spoken hypothetically or corrected an earlier misunderstanding. Those cases need different responses. New location information may supersede old information; a hypothetical statement may never deserve durable memory at all.

A reliable system stores provenance: whether a detail came from an explicit statement, repeated behavior, imported profile or model inference. This makes it possible to resolve conflicts according to evidence rather than just whichever memory retrieval returns first.

Use recency without making it the only rule

The newest message is not always the most reliable. A recent joke should not override an explicitly confirmed birthday, while a direct correction such as “I no longer work there” should have strong authority over an older job record.

Conflict rules can combine source confidence, recency, sensitivity and the type of fact being updated. They should also know when to ask the user a short clarification instead of choosing silently.

Version the active and historical facts

If a city or role changes, the system can mark the previous fact as historical and make the new fact active. This supports continuity without presenting obsolete information as current. The same update should propagate to summaries, embeddings and recommendation features that previously used the old value.

Otherwise, the companion may apologize for a mistake today and repeat it next week because another retrieval layer still contains the older version.

Keep correction controls understandable

Users should be able to inspect important memories and correct or remove them without understanding database architecture. Deletion is different from supersession: someone may want an old job remembered as history, or may want a sensitive fact removed entirely.

Measure conflict resolution longitudinally

Create test histories with repeated preferences, later corrections and conflicting summaries. Score whether the model uses the right active fact, expresses uncertainty where appropriate and respects deletion after several sessions. The benchmark should reward graceful updates, not maximum memory volume.

One practical conflict test uses the same preference expressed in several forms: “I like jazz,” “I used to listen to jazz,” “I’m joking about jazz,” and finally “Please stop recommending jazz.” A naive system can store contradictory strings; a stronger one understands the temporal and conversational meaning behind each statement.

Some conflicts should remain unresolved until the user confirms them. If two reliable sources disagree on a sensitive detail, the companion can avoid acting on that uncertainty rather than choosing whichever memory has the higher embedding similarity. An explicit correction should be easier than forcing the user to restate their history from scratch.

Test whether an update reaches all dependent layers, including session summaries and recommendations. A correct current record is not enough if an outdated embedding continues to surface the old claim several days later.

Conflict resolution rules should consider the type of information. A user’s chosen nickname may remain stable until explicitly changed; a travel date should naturally expire; a medical or highly sensitive inference may not belong in durable memory at all. One universal newest-wins policy cannot capture these differences, so evaluation should be organized by memory category.

Corrections should be idempotent and durable. If a user repeats the same correction, the system should not create several competing versions; if a session is interrupted, the correction should still take effect. Evaluators can test this with repeated edits, device synchronization and a later retrieval call that would previously have surfaced the obsolete fact.