How AI memory handles contradictions and forgetting

An AI memory should treat contradiction and forgetting as different operations. A contradiction says two claims cannot both govern the same scope and time. Forgetting says information should no longer influence recall. Good systems preserve enough history to explain change while keeping stale or unsafe material out of the active answer.

Why append-only memory eventually fails#

Suppose an agent stores:

• “The user prefers detailed answers.”
• “The user now wants concise answers.”

Both statements may be faithful observations. Returning both without time or status forces the model to improvise which one matters. Similarity search cannot reliably solve the conflict because both are equally relevant to the query.

Memory needs state, not only storage.

Four relationships between old and new information#

Update#

The new claim replaces the old claim for the same subject and scope. Preserve the earlier state as history when the reason matters.

Coexistence#

Both claims are true under different conditions. The user may prefer concise status updates and detailed design reviews.

Contradiction#

The claims conflict and the system lacks enough evidence to choose. A human or a trusted rule should resolve it.

Duplication#

The claims express the same fact in different words. They should consolidate without inventing a change.

These cases require more than embedding distance. They depend on entity, scope, time, and source.

Forgetting has several meanings#

“Forget this” can mean:

• stop returning it in normal recall;
• mark it as outdated but preserve history;
• remove it from a derived profile;
• delete it from the underlying store;
• expire it after a retention period;
• revoke one agent’s access while preserving it for another scope.

Products should name these operations precisely. Archiving, invalidating, and deleting have different consequences.

Provenance makes revision safer#

A maintained fact should point to the observation or source from which it was derived. When a contradiction appears, provenance lets a person inspect exact wording, time, and context.

Without sources, the system can only compare summaries produced by earlier models. Each revision then compounds uncertainty.

OceanDB’s maintenance model#

OceanDB separates the immutable observation from the maintained interpretation:

1. Entries record raw facts, decisions, and documents.
2. Pages consolidate current knowledge about an entity or topic.
3. Citation links connect page claims to exact entry quotes.
4. Contradictions and duplicate-page merges become review proposals.
5. Human edits are treated as the new truth and protected from automatic overwrite.

The Dreamer can archive and supersede within its permitted model, but the agent database role cannot delete. This is a deliberate bias toward recoverability.

When deletion is necessary#

Recoverability does not remove legal or personal deletion requirements. A human-controlled system still needs an account-level path for permanent deletion where policy or law requires it. The important boundary is that an autonomous agent should not erase memory merely because a prompt induced it to do so.

Agent retirement and human deletion are different authorities.

Evaluate changing knowledge#

Static question-answer benchmarks miss the hardest part of memory. Add sequences to the evaluation:

1. Record a preference.
2. Correct it with a narrower scope.
3. Ask for the current preference.
4. Ask what changed and why.
5. Archive the old source.
6. Verify that normal recall changes without erasing the audit trail.

Recent memory research increasingly measures update correctness and long-horizon stability alongside retrieval. Are we ready for an agent-native memory system? argues for evaluating memory as an evolving data system rather than a monolithic answer score.

The goal is not perfect remembrance. It is controlled continuity: current enough to act, sourced enough to trust, and recoverable enough to correct.