Actor basis for the write (for example assistant, user, connector, or system). Missing admission stays reviewable.

Hash-only admission envelope. The server emits one stable outcome_class: save, drop, block, or pending_approval. Authoritative admission requires a validated source_event_id and matching workspace; missing or unverified evidence is retained as proposed/requires_review and is not serveable. DROP/BLOCK decisions return hash-only evidence without persisting the candidate.

Agent identity (v1.2.0). Tracks which agent wrote this entity. Used for agent attribution and context filtering.

#849: deliberate trusted override of a rejected-value tombstone. Journaled as an audited override; never set automatically.

JSON object with the entity body — store content, summary, and any custom fields here

Entity category: 'decision', 'architecture', 'convention', 'insight', or custom

#487: the memories this write was built on (max 64). Each cited source is automatically marked useful — usefulness_count bumped, last_useful/last_accessed refreshed — so memories that actually inform later writes rank higher in recall and decay slower. Cite the entities you recalled before composing this write. Unknown citations are reported in the result, not fatal; self-citations are ignored.

Write-time audit envelope for captures and decisions. capture_mode distinguishes snapshot, hash_only, pointer_only, not_requested, capture_failed, and legacy_unknown; a missing value is never interpreted implicitly.

#728: optional first-class pointers to external systems of record (max 32). Stored inside body_json under the reserved 'external_refs' key; filter recall with ref_type/ref_value.

#919: optional 1-3 prospective query hints — natural-language phrasings that should retrieve this entity, indexed into FTS5 alongside the body (vocabulary-gap recall). Default-off: hints are rejected unless the server runs with PERSEUS_VAULT_HINTS_ENABLED=1. Hints replace any previously stored hints on update (omit to clear).

Initial importance 0.0–1.0 — sets the starting decay score

#874: per-write interference bound override — may only TIGHTEN the configured bound (PERSEUS_VAULT_INTERFERENCE_BOUND); a looser bound is refused fail-closed. Writes whose activation overlap with existing memory exceeds the bound are quarantined (default) or refused.

#874: per-write interference-gate mode override. auto (default) uses the operator-configured mode (PERSEUS_VAULT_INTERFERENCE_MODE); refuse/quarantine tighten it per-write. Per-write 'off' is refused fail-closed — only the operator can disable the gate.

Unique key within the category, e.g. 'use-postgres-16' or 'deployment-strategy'

#729: optional memory-origin/provenance metadata (spec: docs/specs/memory-provenance-and-external-refs.md). Stored inside body_json under the reserved 'origin' key — surfaced by recall/get_entity via body expansion. All fields optional; unknown values are left absent, never guessed.

Opt out of near-duplicate merging for this write (#531). Set true for bulk/API ingest of templated records so every acknowledged write actually creates its key; leave false for conversational memory.

#874: sparse update mode — touches only the activated subset of state (body slot, activated links), never disturbs neighbors: no salience inflation on re-assert, caller links admitted only when their target is activated by the new body, no near-duplicate absorption on insert.

Closed lifecycle status vocabulary; proposed/quarantined are never publicly serveable

Tags for categorization and cross-referencing

Hierarchical topic path, e.g. 'architecture/database/postgres'

Entity type: 'insight', 'architecture', 'decision', 'reference', 'convention'

Application-time period start (#363): when the fact became TRUE IN THE WORLD, independent of when it was recorded. Set in the past for retroactive facts ('this was true last week, we just learned it') without rewriting transaction history. Default: transaction time (now). Query with perseus_vault_valid_at / perseus_vault_bitemporal / recall's valid_at filter.

Application-time period end (#363, exclusive): when the fact STOPPED being true in the world. Omit for 'still true' (unbounded). Must be greater than valid_from_unix_ms.

Workspace scope identifier (v1.2.0). Empty = global. Entities with a workspace_hash are invisible to recall queries scoped to a different workspace.

Transport-stamped identity; caller-supplied values are overwritten.

Agent identity filter (v1.2.0). When set, only entities with a matching agent_id are returned. Omit for no agent filtering.

#472 Temporal RAG: transaction-time instant (unix ms). Reconstruct semantic recall AS BELIEVED at this past instant — each hit's body is the version that was live at as_of_unix_ms; corrections recorded later do not leak in. Combine with valid_at for the full bi-temporal cell. Hits are stamped with is_live_version / recorded_at_unix_ms / valid_from_unix_ms / valid_to_unix_ms. Omit for today's live view. (v1: candidate generation is over the live index, so a fact fully deleted since that instant will not surface.)

Filter by category, e.g. 'decision' or 'architecture'

Additive boost for content witness — rewards entities whose body text literally contains query terms. Damped by body length. Never penalizes.

#864: bounded recall. When set, the recall is timed; if it exceeds this many ms the response outcome.status is 'timeout' so callers know the result set may be incomplete. Results are still returned in full.

Fused mode only: depth budget -> default token caps 1024 / 4096 / 16384 when max_tokens is unset.

Per-keyword diversity quota factor (1.0=disabled). Each distinct matched keyword gets ceil(N x halving^n) slots — first keyword N, second N/2, etc.

#880: epistemic trust-axis filter. When set, only entities in the requested trust state are returned — 'candidate' surfaces useful-but-unverified records, 'verified'/'corroborated' restrict to established fact, 'rejected' shows reviewed-and-refused records. Omit for no trust filtering (default).

#1135: opt-in governed answer-facing evidence lanes. Omit for the legacy byte-compatible recall response; choose derived, verbatim, or both under the shared max_tokens budget. Duplicate lane names are canonicalized.

Configuration for FTS5 query expansion using Porter stemming

Fused mode only (#869): graph utility gate threshold in [0,1]. The graph strategy engages only when the query's classified graph utility is >= this value. Omit = 0.5 (documented default). 0.0 disables the gate; 1.0 effectively never engages. The routing decision is always observable in fused_trace.graph_route (reason, selected, skipped_reason, gate counts).

Include archived (soft-deleted) entities in results

Add a normalized confidence score (0.0-1.0) to each result, rolled up from rank, trust (verified/certainty), and decay. Presentation-only; does not change ranking.

#917: add deterministic contradiction/superseded/stale flags containing only entity IDs, validity ranges, and hash-linked claim-card evidence refs. Suppressed values disclose existence only; no body value is rendered.

#917: independently add an ID/hash/validity-only markdown conflict block. Does not implicitly enable structured conflict_flags.

#1142: attach a bounded workspace-scoped hash-only declared graph projection. Requires workspace_hash and a transport-stamped requester; ordinary recall does not query the graph.

#864/#873/#887: always attach the explicit 'outcome' block (status, backend health, abstention, reason). By default it is attached only when recall was degraded/partial/timeout/empty/unavailable/stale, so nominal responses stay byte-identical.

#1141: include only sanitized provider identity, revision, digest, scope, and thread lineage; raw provider bodies and payloads are never returned.

#1140: fused mode only. Attach a bounded, hash-only per-candidate selection projection with source-arm ranks, eligibility/disposition reason codes, token-estimator state, unavailable-arm state, and a replay fingerprint. Omit to preserve the legacy response shape.

Filter by memory layer (world, episodic, semantic).

Maximum number of results to return (max 1000)

Fused mode only: token-budget truncation (estimated tokens = chars/4 per body). 0 = derive from depth_budget (mid = 4096).

Minimum decay score threshold 0.0–1.0 — higher values return fresher results

Search mode: 'fts5' (keyword), 'dense' (vector), 'hybrid' (fused via RRF), or 'fused' (TEMPR-style multi-strategy: fts5 + dense + graph + temporal with weighted RRF, token-budget truncation, and a full fused_trace, #883)

Number of results to skip for pagination

If set, truncate body_json at N chars and append drill-down footer. Use perseus_vault_get_entity to read full body.

#860: validity-aware recall profile. 'validity' re-ranks fused results by a deterministic validity multiplier (freshness decay, scope match, provenance class, supersession, expiry proximity) and annotates every item with its validity info; 'default'/omitted keeps relevance-only ordering. On non-fused modes the profile only enables item annotation. The weights, grade distribution, and context-invalid count are observable in fused_trace.validity.

Search query — words are OR'd together for broad recall. An EMPTY string ("") is the match-all / enumeration path: it drops the keyword predicate and returns every entity in scope (respecting category/type/limit/offset), so it is the way to 'list all' a category. Wildcards are NOT globs: "" is a literal FTS5 term and matches nothing — pass "" to enumerate, not "".

Fused mode only: anchor instant for the temporal strategy (unix ms; default now). Accepts a number or numeric string.

Time-aware ranking for mode='hybrid' (default off). When set, each fused result's score is multiplied by 0.5^(age / this), where age is seconds since the memory was created — so a memory this many seconds old keeps half its weight and recent context outranks older but similar hits. Omit for relevance-only ranking.

#728: post-filter hits to entities whose body external_refs carry this ref_type (exact match, e.g. 'repo', 'pull_request', 'jira_key').

#728: post-filter hits to entities whose body external_refs carry this ref_value. Matches exactly or as a hierarchical '/' prefix ('github:Org' matches 'github:Org/repo').

Opt-in reinforcement for mode='dense'/'hybrid': bump retrieval_count/last_accessed/decay on the returned hits so semantically-used memories resist decay and promote through layers. Default false keeps semantic recall side-effect-free and byte-deterministic over a frozen DB. No effect on mode='fts5', which already reinforces.

Transport-stamped requester identity used for private/fleet visibility enforcement.

Fused mode only: optional rerank stage over the fused pool (rank-calibrated dense + BM25 agreement signals; default off, latency-preserving).

#784 serving posture. personal returns preference/personal classes; agent returns convention/correction/keystone classes; shared (default) returns non-personal memory in the requested workspace. Applied after visibility filtering.

#485: scope as a ranking multiplier instead of a hard filter. Requires workspace_hash. Widens the workspace filter to also include GLOBAL (workspace_hash='') memories, weighted by this factor in the ranking (hybrid/dense scores multiplied; keyword mode returns current-scope hits first) — current-workspace memories outrank equally-relevant global ones, but a strong global memory still surfaces. Never exposes other workspaces' memories. Omit for the strict filter (unchanged default).

Fused mode only: strategies to engage (2-4). Omit = all four. Unknown names are rejected.

Fused mode only: per-strategy RRF weight multipliers (default 1.0 each). Arms that find nothing contribute nothing.

Filter by topic path prefix, e.g. 'architecture/'

Additive boost for provenance/trust (default 0.15, on by default) — verified sources rank above unverified AI drafts on the same topic. Verified entities get the full boost; unverified ones are scaled by certainty. Set 0 to disable. Never penalizes.

Filter by entity type, e.g. 'insight' or 'reference'

Valid-time instant (#363/#472, unix ms): reconstruct recall to the world-version whose application-time period [valid_from, valid_to) contains this instant — 'what was true at time T', per current (or as_of) knowledge. Rebuilds the point-in-time body from history (not just a live-row narrow) and returns hits stamped with is_live_version / recorded_at_unix_ms / valid_from/to. Combine with as_of_unix_ms for the full bi-temporal cell.

Valid-time period filter start (#363, unix ms). Pair with valid_to_unix_ms and valid_op; ignored when valid_at is set. Omit for unbounded start.

SQL:2011 period predicate for the valid-time period filter (#363): 'overlaps' (fact's valid period shares at least one instant with the queried period) or 'contains' (fact's valid period contains the whole queried period).

Valid-time period filter end (#363, unix ms, exclusive). Omit for unbounded end.

#860: annotate delivered items with their validity info (grade, freshness, scope match, provenance class, superseded, expiring/expired, multiplier, signals); context-invalid items are additionally flagged 'context_invalid': true. Implied by profile='validity'.

Workspace scope filter (v1.2.0). When set, only entities with a matching workspace_hash are returned. Omit for no workspace filtering.

Transport-stamped requester identity applied to every nested query and fused result.