Entity Resolution for Agent Memory: Avoid Connecting the Wrong Facts

Entity resolution determines whether two references describe the same person, organization, project, or other subject. In a memory graph, a mistaken merge can attach correct facts to the wrong entity and contaminate many later answers. Resolve identity with provenance and uncertainty rather than relying on similar names alone.

Separate labels from identity

Two employees can share a name, a project can be renamed, and one organization can have several aliases. Keep stable source identifiers where available and preserve the system they came from. A display name is useful for reading but usually insufficient as the only identity key.

Do not merge across tenant or workspace boundaries just because names and descriptions match. Authorization scope should constrain candidate matches before semantic similarity is considered.

Record why a match was made

Useful evidence may include a shared verified identifier, an explicit source statement, or an application-owned mapping. An inferred match should remain distinguishable from a confirmed one. Store enough provenance to revisit the decision when new evidence appears.

Choose conservative behavior for ambiguous cases. Keeping two unresolved candidates can be preferable to confidently combining them. The application can ask for clarification or retrieve source-specific evidence without asserting an identity match.

Plan for splits and corrections

A graph must be able to recover when an entity was merged incorrectly. Identify which relationships and derived memories came from each source so they can be reassigned or invalidated. A simple rename is different from a split: changing the label does not repair facts attached to the wrong subject.

Test an organization rename, two people with the same name, a reused email address, and contradictory source records. Inspect downstream answers after correcting the identity mapping, not only the graph's visual appearance.

Evaluate the questions affected by identity

Use questions that require connecting records across sources and questions that should remain separate. Measure incorrect merges and missed matches independently; they have different consequences. Include access changes and deletion in the lifecycle test.

The relationship-semantics guide explains how relationships affect memory. To understand a managed implementation before evaluating it, read Supermemory's graph-memory concepts and test your own ambiguous-identity examples against the documented behavior.