Entity identity and conservative merges

A memory owner is the account storing a message. A message role identifies its source, and an entity identifies the subject of a claim. These are separate: a user can paste a colleague’s email without adopting the colleague’s preferences.

Extracted English personal references such as I, we and they remain local to their source event. They carry identity_status="unresolved_reference" and reference_event_id metadata. Repeating preparation for that same event may reuse its reference; another event, absent provenance, or an older globally merged pronoun node cannot establish the same referent. Explicit non-personal meanings such as US/country and IT/technology retain named-entity behavior. This is a conservative English guard, not general coreference resolution. It does not disambiguate several quoted speakers inside one event or equal names belonging to different people.

enable_speaker_references (on by default since 2026-10-06) provides a named exception. When a turn names its speaker (ingest(speaker=...)), a first-person singular reference in its claims (I, me, my, mine, myself) binds to an entity with that speaker’s name and type person, the same node that other mentions of the name use. The claim keeps the literal pronoun as its subject or object, and its speaker_reference metadata names the speaker and which of the two it bound. A speaker’s claims then connect across messages, so a repeated claim can merge (#209) and a named update can supersede the value it replaces. Turns without a speaker or an explicit owner declaration, and plural, second-person and third-person references, stay local to their message. It became a default together with enable_claim_sentence_text and packing.fold_repeated_text, after the three passed the default-change rule on the DeepSeek answer track over ingest() packs (BENCHMARKS.md, “Default change: claim sentences, speaker references and folding”). Only the caller knows that the named speaker, and not someone the turn quotes, is who says I, so set PRME_ENABLE_SPEAKER_REFERENCES=false when a speaker’s turns quote others in the first person.

With enable_ingest_claim_fixes (default true since 2026-10-19 in the alpha channel; false restores 0.16.0a2), a user turn that gives no speaker but says who it is (“Hi, I am Alex Rivera”, “I’m Alex”, “My name is Alex Rivera”, “Call me Alex”) binds the same way when the extraction lists that name as a person; the claim’s speaker_reference metadata then has "basis": "self_introduction". That binding is local to the turn that names itself, unless enable_owner_name_carry (below) is on, as it is by default: later turns of the session still keep their own I, a caller’s speaker or first_person_owner wins, and a possessive (“I am Jordan’s partner”), a name the extraction did not list as a person, and a turn that introduces two different people bind nothing. The default rests on the fast gate by the owner’s directive of 2026-10-19; the full benchmark audit and the default-change rule on the ingest() track are pending.

With enable_owner_name_carry (default true since 2026-10-19 in the alpha channel; false restores 0.16.0a2; it needs enable_speaker_references), the name outlives the turn that gave it. A user turn is the memory’s owner by definition, so the person entity that a user turn creates under the owner’s name (the caller’s speaker, or the introduction above) carries owner_name metadata, written once with the node. A later user turn of the same owner and scope that names nobody (“I prefer trains to flights”) binds its I, me and my to the newest such entity, so the person is linked to all of their claims and no I entity is made for the message; the claim’s speaker_reference metadata then has "basis": "owner_name". It never applies to participant, assistant, tool or system turns; a caller’s speaker or first_person_owner on the turn wins; a later name replaces an earlier one for the turns after it and earlier claims keep theirs; and a name no user turn gave is never invented. The extraction request, the number of model calls and the claim wording are unchanged: a statement written for a duplicate sentence still says “the speaker”, because a name in every such statement makes a question that names the owner match all of them. Only an entity that the turn creates is marked, so a person an earlier message already named is not, and an owner who goes back to an earlier name (A, then B, then A again) keeps B, because A’s entity is not marked again. A failed lookup leaves the old behavior. The owner’s entity lives in one scope, so a name given in one scope is not carried to another.

With enable_owner_group_links (default true since 2026-10-19 in the alpha channel; false restores 0.16.0a3; it needs enable_owner_name_carry), a claim that a user turn states about we, us or our is linked to the people it speaks for, while its subject stays the we entity local to its message and its text keeps the literal word (prism-monorepo issue 275). We always includes the speaker, so the claim gains a MENTIONS link marked metadata.basis group_member to the owner’s named entity (the newest entity a user turn marked with the owner’s name, or the one under the turn’s own speaker). It is linked to the other members only on evidence: when the session’s user turns with the same speaker, up to and including the claim’s turn in the order they were ingested, name exactly one group with the owner (“My partner Jordan and I”, “Jordan, Sam and I”, “me and Jordan”), the claim is also linked to each named person, and that link records the turn that named the group in named_in. A member must be a person entity a turn already stored, matched by its full name (or, after a weekday, a month or a title such as “Sunday Jordan” or “Coach Taylor”, by the rest of it), and only one person may carry that name, so “Jordan Lee and I” never links a stored “Lee”. A group named anywhere earlier in the session counts as evidence, however far back and whatever it was about: in a long session, a later “we” about something else (a work decision after “Jordan and I went hiking”) is linked to Jordan too. Only the newest 500 events of the session up to the claim’s turn are read, only the first 4,096 words and marks of a turn, and a list of more than 16 members names no group. When the session names two or more different groups, a member without a name (“my wife and I”), or a name no single person entity has, the claim is linked to the owner only; no member is invented, and names count as written with a capital letter. The reading is a conservative English rule, not coreference resolution: “and I” counts only after members that start a clause, so “I got the offer from Brightpath and I accepted it” names no group. Participant, assistant, tool and system turns keep their own we, as do first_person_owner turns. With enable_claim_merge, the same we claim repeated in two user turns of one session merges into one current record, as a named speaker’s repeated I claim does: the earlier copy is found through its link to the owner who said it, so another speaker’s identical claim, linked to the owner only as a member of that speaker’s group, never retires it. A claim repeated in another session stays separate, because a session may speak for another group. A claim whose sentence names the owner still gets its group link beside the text link, so it can merge too. The statements built for a duplicate sentence keep “the speaker’s group” and never name the members. The links are made after the derivation commits, outside the plan, so saved plans keep their bytes; memories ingested before the option was on are not linked or merged. reingest() makes the same links and merges, with one gap: it re-derives a session in event-time order while the links follow ingestion order, so when a session’s turns were ingested out of event-time order, a repeated we claim can be re-derived before its earlier copy has a link, and that merge is missed. The default rests on the fast tiers by the owner’s directive of 2026-10-19; the full benchmark audit and the default-change rule on the ingest() track are pending.

ingest(first_person_owner=True) provides a separate per-message declaration for user/human sources with no named speaker. Under fresh v14 plans, singular first-person references bind to one deterministic owner_reference entity per owner and scope. Its content remains I; it binds no display name. Claims keep their literal references and record owner_reference with the owner and bound fields. The reserved source metadata key is prme_first_person_owner_v1. This is caller-declared authorship, not automatic coreference. Leave it false for pasted first-person letters; split mixed authorship into separate messages. Old unresolved identities and named speaker entities are not merged into this identity. See the correction API.

enable_speaker_grounding (default false) makes the same assumption during grounding. Grounding requires a claim’s subject and object to occur in its quoted sentences, and an extractor writes the speaker into a claim as their name (“Dana, lives in, Denver” for “I live in Denver”) or as I (“Rachel, is sister of, I” for “my sister Rachel”), so grounding discards a name absent from its evidence. Fresh extraction also accepts an unnamed first-person singular subject (I, me, my, mine, myself) when any of those forms occurs in its cited sentence. For example, I / name / Sam is supported by “My name is Sam”, and a preference I / wants_to_be_called / Sam by “Call me Sam”. This binds no personal name and keeps the reference event-local. It does not infer a missing first-person subject in “keep answers short”, bind plural references, or establish that pasted first-person text belongs to the memory owner. With the option, in a turn that names its speaker, the speaker’s own first-person singular reference (I, me, my, mine, myself) counts as a mention of a claim value that is the speaker’s name or one of those references. That applies in the provider’s check, the pipeline’s second check and the sentence choice of enable_claim_sentence_text. The model is asked the same question either way. A first-person attempt or intention (“I’m trying to set up ESLint”) still needs a predicate that keeps it when the claim names the speaker, and a claim naming the speaker covers the speaker’s I clause, so recovery adds no duplicate for it. Plural references are unchanged.

Built-in extraction keeps only claims whose named references resolve within the same response, apart from those event-local personal references. A missing or ambiguous reference discards its claim while preserving grounded, closed siblings; it never creates or selects an identity to repair model output.

enable_pronoun_repair (on by default since 2026-10-22 in the alpha channel; dwamianm/prism#345 and #347) names the person behind a third person pronoun that a first extraction left as a claim’s subject or object, for a memory that names no one (“They chose Tokyo, Kyoto, Kanazawa, and Osaka.”). It asks one more question, only for a source with such a claim: the owner’s preceding memories (up to extraction_window_turns, in the order stored, in the same scope, or the turns before it in its session) and only that claim’s own sentences go to the model, which returns names or nothing. It is not coreference resolution. A first person pronoun is never repaired: I is bound by the speaker rules, and we, us and our include the speaker and a group no memory states, which the owner group links connect to the people a session names with the owner. The first variant also repaired we, and turned “We use Tableau at Northwind.” into “Dana and Priya use Tableau at Northwind.”, a membership nothing stated.

Grounding requires a claim’s subject and object to occur in its quote, and a repaired name does not occur there. The quote stays the claim’s evidence, verbatim, and its check is unchanged. The name has its own rule: it must be a proper name (a capital letter, not a pronoun) written in the claim’s sentence, the message, the speaker line, or a memory the request showed, and when a pronoun resolves to several names every one must pass or the field is left as it was. A claim is repaired only in its own fields; no other claim of the message changes, and a failed or refused answer leaves the first extraction. A repaired claim’s text is its own sentence with the pronoun replaced by the names (grounding_method pronoun_repair_v1). A field that resolves to one name is rebound to that person’s entity, so the claim attaches to them. A field that resolves to several (they) stays on the pronoun entity, its text names every member, and each name is listed as an entity so the mention links reach each one.

A repaired claim records what it used in metadata.resolved_from_context: {"event_ids": [...], "fields": {"subject": {"pronoun": "They", "names": ["Alex", "Jordan"]}}}. event_ids are the memories of the request’s window whose text wrote any of the names (empty when the names were in the claim’s own message), and fields has an entry for each repaired field, subject or object. metadata.resolution holds the sentence it replaced, each replacement and where each name was found, and metadata.evidence_quote is the original quote. The repair runs before the extraction is saved, so reingest() replaces the pronoun version like any other derived claim. With the option off (PRME_ENABLE_PRONOUN_REPAIR=false) no request is made and a saved extraction has no new field. It was turned on by the maintainer’s decision, without the alpha pass rule’s defect probe condition met and without a mid tier or full benchmark run on it (BENCHMARKS.md, “Pronoun repair”).

Fresh extractions and plans record speech_act_v14. V14 preserves v13 unnamed singular-form matching and earlier qualifier, quantity, speech-act and source-effective validity rules, and adds explicit owner references and bounded correction matching. Saved v13 records still prepare v13 plans. Saved v11 and v12 extractions keep their v12 plan behavior; saved v12 plans retain literal-only unnamed matching. Saved v10 through v6 records also prepare v12, v5 records prepare v11, v4 prepare v10, v3 prepare v9, and v2 prepare v8. A v12 extraction record also prepares v12. Legacy source_passage_v1 extraction records prepare missing plans under temporal_validity_v7, so recovery never claims validation that did not run. Prepared plans using speech_act_v11, speech_act_v10, speech_act_v9, speech_act_v8, temporal_validity_v7, grounded_quantities_v6, claim_qualifiers_v5, event_local_references_v4, and earlier policies remain valid immutable replay inputs. Previously saved plans preserve their policy, node IDs and checksums and replay unchanged. The default for an omitted historical policy remains the previous value. Existing incorrectly merged identities are not automatically split; doing so requires an explicit provenance-based correction.

Organizer duplicate and alias candidates must agree on owner, scope, memory type, source type, epistemic classification, session, event time, validity end, retention settings, pinning and metadata. This includes entity types and claim subject/predicate metadata. Unresolved personal references cannot become canonical identities through organizer merging or alias links.

For non-entity duplicate copies, automatic merging additionally requires exactly equal content and validity start. Two separately admitted observations remain distinct even when their text matches. Case changes in ordinary claim text are not assumed equivalent. Named entities retain stripped, case-insensitive name matching within the compatible metadata/provenance boundary. This policy favors retaining an uncertain duplicate over erasing an episode or qualifier.

Vector similarity still proposes candidates; it does not authorize replacing different text with a canonical claim. organize(jobs=["deduplicate"]) reports both candidates and actual merges, plus unapplied pairs and its merge-policy version. A candidate count greater than zero with no merged nodes can be correct. Compatible exact copies can still merge and combine their evidence references.

Organizer duplicate and alias merges now publish canonical evidence, copied relationships, retirement and one supersedence edge in a single backend transaction. They revalidate both nodes after entering that transaction. PostgreSQL locks shared nodes in UUID order before reading evidence, so overlapping merges cannot overwrite each other’s evidence union. A conflicting DuckDB transaction can fail safely and be retried against current state.

Copied relationships retain validity, provenance, confidence, metadata and assertion time. Their deterministic IDs also recognize exact partial copies left by older versions. A conflicting existing copy aborts the new transaction. Each committed merge records complete before/after node values, original and published relationships, and a checksum-protected ORGANIZER_MERGED operation. The operation identity is stable for the unordered node pair and merge kind. Non-finite numeric metadata is rejected before commit rather than silently changed to JSON null in the journal. Existing compact JSON records retain their original bytes, checksums and retry identities. Retries return the committed identity without reactivating later-retired nodes.

External index eviction runs after graph commit and remains repairable by compaction. Retired graph state governs candidate admission if eviction fails. Cancellation or a lost acknowledgment may follow a committed transaction; callers must not interpret cancellation as proof of rollback. This adds complete records for new organizer merges, not replay of every historical manual or organizer mutation.

Alias application rechecks the actual names. A compatible known abbreviation, case variant or exact normalized name can merge at the configured confidence threshold. A high semantic score by itself only creates a RELATES_TO proposal marked identity_verified=False, preserving both entities. Caller-supplied candidate labels do not bypass these checks. A proposed alias is not a verified same-person relationship.

New proposals publish their relationship and a checksum-protected ALIAS_PROPOSED record in one backend transaction. Their operation and edge IDs are stable for the unordered pair, so repeated or concurrent organizer passes do not create duplicate links. The record retains both complete node inputs and the exact edge. Version-2 records can additionally retain a complete external assessment bound to exact node checksums. Repeated matching requests return the first committed assessment; older proposals are never relabeled with evidence they did not originally contain. Existing random-ID proposal edges are recognized and reused but not backfilled with invented creation history. Unverified proposals remain queryable as edges but are excluded from ordinary neighborhood and shortest-path traversal unless a caller explicitly opts in.

An owner can list these records through the alias-proposal inbox and make one explicit accepted or rejected decision. The decision transaction retains the complete original proposal payload and reviewer input in a checksummed record. Acceptance rechecks the exact proposed node snapshots and publishes a separate verified alias edge; it leaves both nodes active and does not perform an organizer merge. Rejection publishes no traversable edge. Matching retries reuse the deterministic review record, while conflicting decisions fail. This turns a model or similarity signal into a reviewable identity link without treating it as permission to delete or supersede either entity.

These guards protect source identity and merge behavior. They do not establish that extraction predicates are entailed, that a source is truthful, or that a reader correctly attributes every quotation. The PersonaMem reader diagnostic still exposes attribution failures on raw histories; it is not an evaluation of these new ingestion/organizer guards.

Source: docs/ENTITY-IDENTITY.md  in PRME at bd92166f.