Identity and Relationships

Two statements can look alike without describing the same entity or making the same claim. A person can share a name with someone else; the same person can be described in different words; and claims with similar wording can differ in time, scope, perspective, or polarity. The maintained-memory plan separates candidate matching, identity decisions, and graph relationships so one signal does not silently decide another.

Matching proposes; rules decide

Vector similarity and keyword search can find candidate memories for review. They cannot establish that two claims mean the same thing, that two people are identical, or that two sources independently corroborate one another. Likewise, different wording does not prove that claims differ.

The planned exact identity check for structured claims considers the predicate, entities, typed value, polarity, scope, perspective, and applicability. A repeated text hash is useful only when its surrounding context also matches. Ambiguous semantic interpretations remain separate until the conditional acceptance rules resolve them. A matching process must also account for concurrent changes: if another claim appears after candidates were gathered, the decision must be checked against the newer state before it is accepted.

Matching and relationship discovery solve different problems. Matching can attach new evidence to an existing exact claim rather than creating a duplicate. Discovery can still find a meaningful relationship between two distinct memories; one pipeline does not replace the other.

Merge and split require identity evidence

A future merge or split is a versioned decision with explicit parents, lineage, aliases, and dependent contributions. The plan requires corroborated identity evidence, such as matching persistent external identifiers or explicit identity reports. Similarity by itself is not enough.

Original memory-version references remain intact. A merge does not rewrite all older evidence to point at one surviving memory. A split attributes evidence, usage, and graph links to the versions they came from rather than copying all support to every child. If an identity decision is later withdrawn, dependent claims are reconsidered while independent confirmations can survive.

This keeps identity uncertainty visible. When the evidence does not establish whether two records refer to the same subject, keeping them separate is safer than making an irreversible-looking merge that later claims inherit.

Graph links carry provenance, not belief

The relationship graph describes connections such as related-to, caused-by, part-of, references, supports, contradicts, or derived-from. Planned graph provenance records whether a link was asserted, discovered, or derived, the decision that created it, and the exact memory versions it connects.

A relationship named supports is still a graph interpretation. It does not become an eligible belief justification, registered derivation, or automatic withdrawal of another claim. Belief changes require their own accepted contribution under the belief rules. In particular, a model’s causal link is not proof that one event caused another.

Imported links keep whatever origin information is actually known. The system must not guess an unknown origin, turn an observation reference into an arbitrary focused memory, or treat an imported supports edge as evidence. Adopting an imported link into the managed graph requires explicit endpoint mappings to exact memory versions.

Typed relationship discovery

The plan proposes extending the existing model-free similarity discovery with an optional model-backed step. It would retrieve bounded candidates and ask a configured model to propose one of the already registered relationship types, with direction and bounded rationale/evidence references, or no link. The model cannot invent a relationship type. Proposals are checked against tenant, version, evidence-span, lifecycle, privacy, revision, and processing-generation boundaries before they can change the graph.

The relationship strength is a graph-ranking measure, not a probability that the relationship is true. A generated link records its provenance so it can be reevaluated or deactivated if its source version changes. Caller-asserted and independent links must not be overwritten when one model-generated link is withdrawn.

The plan calls for multilingual fixtures that distinguish relatedness, directional causality, part-of, reference, support, contradiction, and no-link cases. High similarity alone must not determine the type, and structural validation cannot substitute for measuring whether the model interpreted the relationship correctly.

For belief grades and the limits of support, see Beliefs and Evidence. For broader patterns inferred from several memories, see Consolidation.