Recall Ranking

Recall Ranking

Recall returns authorized evidence contexts in results, plus separate Relationship, community, conflict, and Hypothesis fields. Stored evidence must pass final active-support and time-view eligibility before entering primary results.

Read Shapes

Operation	Result
recall_memory	Ranked evidence contexts, related frontiers, conflicts, search states, degradations, and suggested actions.
trace_memory	One authorized Relationship's bounded evidence, support decisions, transitions, and graph/conflict lineage.

Array order conveys rank. Public results do not expose raw vector distances, fusion scores, or provider attempts. Exact schemas are in Technical Reference.

Default Pipeline

%%{init: {"flowchart": {"curve": "linear"}}}%%
flowchart LR
  Request["Authenticated query and time view"] --> Scope["Authorized memory-space branches"]
  Scope --> Evidence["Full-text, vector, and optional Entity expansion"]
  Evidence --> Fuse["Evidence-ID rank fusion"]
  Fuse --> Hydrate["Canonical source/support/time hydration"]
  Hydrate --> Results["results"]
  Scope --> Related["Shared Relationship discovery"]
  Related --> Coverage["Exclude covered semantic groups"]
  Results --> Coverage
  Coverage --> Community["Current communities with novel members"]
  Coverage --> Direct["Disjoint direct Relationship fallback"]
  Scope --> Hypothesis["Optional shared Hypotheses"]

Loading

A query embedding is reused across allowed-space branches. Candidate queries retain team and space scope before IDs leave the adapter. Search text never replaces canonical evidence hydration.

Eligibility

Evidence needs an eligible active Relationship support path for the requested view. Hydration checks evidence/source lifecycle, support decisions, Relationship status, authorization, memory-space generation, and conflict disposition.

valid_at selects real-world time; known_at selects system knowledge time. A later source withdrawal or conflict resolution must not rewrite what an authorized earlier knowledge view reports.

Safe evidence stored without a Relationship can have a current search document while remaining absent from default recall. Candidates and Hypotheses do not satisfy the support requirement.

Evidence Frontier

The evidence adapter combines full-text and vector ranks with weight 1 each, and explicit Entity expansion with weight 0.5, using RRF constant 60. Known evidence IDs are excluded. Stable equal-score order uses evidence ID.

A bounded overfetch supplies candidates for canonical hydration. Ineligible hits are discarded; errors loading authoritative state fail visibly rather than becoming an empty result.

Relationship Discovery

Shared Relationship discovery retrieves active supported candidates from lexical and eligible vector branches with equal-weight RRF. Equivalent authored Relationships collapse by semantic group without merging ownership. Returned handles retain representative and equivalent IDs.

Current and known evidence plus known Relationship IDs establish covered semantic groups. Later related frontiers suppress those groups to avoid repeating the same supported meaning.

expand_from_entity_ids requests bounded direct expansion, not unbounded multi-hop traversal. Source authority and supporter count are provenance and conflict-policy inputs, not an arbitrary popularity boost to query ranking.

Community Frontier

Community selection uses the current authorized snapshot and uncovered semantic groups. Returned communities contain novel Relationship members. Emitted community groups are removed from the direct fallback; singletons or members omitted by community limits can remain eligible for that fallback.

Community output has an independent budget and does not replace evidence results. Temporal recall skips communities with an explicit degradation. Missing, stale, failed, or oversized enabled snapshots also degrade explicitly.

See Community Memory.

Conflict Disposition

Relationship conflict cases retain authored alternatives and supporter summaries. Resolved preference affects eligible current support paths; temporal views reconstruct the applicable earlier case state.

Evidence conflicts are separate kind: evidence_conflict entries with exact occurrence citations. Only conflicts reachable from authorized evidence are projected. Case preference does not turn a quote into independently submitted evidence. See Conflicts And Corrections.

Independent Space Branches And Fusion

A private-bound principal can read shared memory and its permitted private branch. Evidence search and hydration run independently in each space. The service fuses authorized results with equal-weight RRF constant 60; stable ties prefer private, then shared, then immutable ID.

The current related Relationship/community/Hypothesis expansion is restricted to the shared branch. Do not infer a full private graph frontier from the fact that private evidence is searchable.

Results expose space_kind, not an internal space UUID. Neither Entity resolution nor graph expansion joins the shared and private branches.

Failure, Degradation, And Freshness

Required repository or hydration failures remain errors. Optional vector or derived-frontier loss is reported in degradations with bounded code, frontier, and message. Inspect these fields before describing a result as a complete search.

Relationship vector readiness is gated by its projection generation. Pending or failed vectors can leave valid lexical discovery available with explicit state and degradation. Successful Remember writes have already completed their required embeddings; this is not a post-response Remember queue.

Recall-feedback snapshot failure preserves usable recall but omits its feedback suggestion and reports recall_feedback_snapshot_unavailable.

Trace And Suggested Actions

Use trace_memory for support, decisions, and lifecycle details of a returned Relationship. Respect its bounds and stopped_reason. A partial lineage view cannot establish a complete correction support set.

When feedback is enabled and its snapshot persists, suggested_actions provides the recall event ID for submit_recall_session_feedback. Effective Dreaming can suggest resolve_dream_feedback for returned Hypotheses. Feedback records use; it does not change ranking or lifecycle.

Risks And Verification

Test stale search hits, revoked support, known-equivalent suppression, community/direct disjointness, temporal conflict views, private/shared fusion, and required/optional failures. Include different owners and another team in real PostgreSQL cases.

All returned text remains untrusted. The host's action and disclosure policy must not be replaced by instructions embedded in recalled evidence.