Memory Recall Event
Describe one agent-memory retrieval attempt: its query, authorized scope, observed candidates and ranking, and evidence of downstream selection or use.
Bundle → Layer → Finding → Questions Filled
6 bundles · 12 layers · 12 findings · 48 questions
Recall request Occurrence identity and submitted retrieval intent.
Recall occurrence identity
One bounded agent-memory retrieval attempt is an event even when it runs for an interval. A repeated execution gets a new attempt identity; redelivery and corrections refer to the original occurrence. The recorded trace is an artifact of that occurrence.
Recall occurrence identity
One bounded agent-memory retrieval attempt is an event even when it runs for an interval. A repeated execution gets a new attempt identity; redelivery and corrections refer to the original occurrence. The recorded trace is an artifact of that occurrence.
- Which master identifier and producer scope identify this recall attempt? identity
- Is this a new attempt, a transport duplicate, or a correction of an existing event? classification
- Which request, session and optional retry chain connect the event to its caller? relationship
- When did the attempt start and end, and when was its evidence observed? temporal
Query and retrieval intent
Capture the effective retrieval request and its declared transformation path. Text, identifier, vector or mixed inputs are possible; opaque or redacted inputs are explicit. No confidential query or embedding is required in general logs.
Query and retrieval intent
Capture the effective retrieval request and its declared transformation path. Text, identifier, vector or mixed inputs are possible; opaque or redacted inputs are explicit. No confidential query or embedding is required in general logs.
- What retrieval purpose and input form were declared by the caller? definition
- How does the effective query differ from the submitted query? provenance
- Which requested limits, filters and ranking profile constrained the attempt? constraint
- Which query fields may be retained and which must be minimized or withheld? privacy
Authorized search context Recorded authority and observed searchable state.
Scope and authority evidence
Record the caller, tenant and purpose scope plus externally issued authorization decisions for this attempt. Policy and grant masters decide eligibility. A retrieval result neither grants access nor licenses subsequent use.
Scope and authority evidence
Record the caller, tenant and purpose scope plus externally issued authorization decisions for this attempt. Policy and grant masters decide eligibility. A retrieval result neither grants access nor licenses subsequent use.
- Whose caller and tenant scope applied when retrieval was requested? ownership
- Which policy and grant revisions authorized, denied or left this attempt unresolved? authority
- At which retrieval and handoff stages were authorization checks evidenced? access
- How are denied identifiers and counts concealed from recipients without disclosure authority? security
Store and index observation
Reference each queried store and index revision, observation boundary and freshness claim. Index refresh, memory validity and record trust remain external. Reproducibility is qualified when the snapshot or implementation state is unavailable.
Store and index observation
Reference each queried store and index revision, observation boundary and freshness claim. Index refresh, memory validity and record trust remain external. Reproducibility is qualified when the snapshot or implementation state is unavailable.
- Which stores, indexes and namespaces were actually queried? composition
- Which versions or observation boundaries identify the searched state? provenance
- What freshness or validity constraints were requested and observed? temporal
- Which dependencies prevent another reviewer from reproducing this attempt? quality
Observed execution Stage evidence and qualified completion.
Observed retrieval stages
Record the observed path through retrieval, merge, filtering and optional reranking. Stage identity preserves branch and retry attribution. This model records runtime evidence; it does not execute search, build indexes or implement ranking.
Observed retrieval stages
Record the observed path through retrieval, merge, filtering and optional reranking. Stage identity preserves branch and retry attribution. This model records runtime evidence; it does not execute search, build indexes or implement ranking.
- Which retrieval, merge and reranking stages were observed for this attempt? process
- Which implementation and parameter revisions were reported for each stage? provenance
- How were parallel branches and cached results linked to their original evidence? relationship
- What latency and resource observations are available with units and measurement basis? measurement
Completion and result completeness
Completion, coverage and delivered cardinality are separate observations. Empty, denied, failed, cancelled, timed out, partial and unobserved outcomes remain distinguishable. A successful transport response is not proof of exhaustive search.
Completion and result completeness
Completion, coverage and delivered cardinality are separate observations. Empty, denied, failed, cancelled, timed out, partial and unobserved outcomes remain distinguishable. A successful transport response is not proof of exhaustive search.
- What completion state and terminal evidence were reported for the attempt? state
- Was the result complete within the declared scope, partial, or of unknown completeness? quality
- Are reported match counts exact, lower bounds, estimates, withheld or unknown? measurement
- How are timeout, cancellation and failure distinguished from a verified empty result? exception
Returned results Candidate identity and stage-specific ordering.
Returned candidate identity and lineage
Candidate membership identifies a specific returned memory item or passage representation at an observed stage. Store and version scope prevent collision. Passage selectors locate content; snapshots do not duplicate authoritative memory ownership.
Returned candidate identity and lineage
Candidate membership identifies a specific returned memory item or passage representation at an observed stage. Store and version scope prevent collision. Passage selectors locate content; snapshots do not duplicate authoritative memory ownership.
- Which candidate entries were returned at each observed stage? identity
- Which item or passage version and locator identify the returned representation? provenance
- How are repeated passages and multiple representations linked without collapsing their lineage? relationship
- What does a missing candidate payload mean for inspection of this trace? evidence
Rank and score interpretation
Keep stage-specific rank, score scale, ordering direction and tie behavior. Raw similarity, distance, reranker values and fused ranks are not interchangeable. A high rank is not a truth, permission, trust or calibrated confidence judgment.
Rank and score interpretation
Keep stage-specific rank, score scale, ordering direction and tie behavior. Raw similarity, distance, reranker values and fused ranks are not interchangeable. A high rank is not a truth, permission, trust or calibrated confidence judgment.
- What order and tie policy were reported for the returned candidates? classification
- What does each score measure and in which direction is it ordered? measurement
- How did reranking or fusion change candidate order between stages? process
- Where are score values missing, incomparable or unsupported by an explanation? quality
Selection and use evidence Separate handoff records from attributed downstream use.
Selection and downstream handoff
Record selection or exclusion from the observed candidate set and receipts of downstream handoff. Context Assembly owns packing, budget allocation and actual context membership. Selection, delivery and acknowledgement are distinct from use.
Selection and downstream handoff
Record selection or exclusion from the observed candidate set and receipts of downstream handoff. Context Assembly owns packing, budget allocation and actual context membership. Selection, delivery and acknowledgement are distinct from use.
- Which returned entries were selected or excluded for downstream handoff? decision
- What externally reported reasons account for selection or exclusion? evidence
- Which Context Assembly or consumer revision acknowledged which candidate representations? relationship
- Was access eligibility rechecked when cached or delayed results were handed off? authority
Evidenced downstream use
Record attributed use assertions with their evidence and destination revision. Context inclusion, an emitted citation and assessed support are different claims. Answer Grounding owns statement-to-source support assessment; recall cannot infer private model reliance.
Evidenced downstream use
Record attributed use assertions with their evidence and destination revision. Context inclusion, an emitted citation and assessed support are different claims. Answer Grounding owns statement-to-source support assessment; recall cannot infer private model reliance.
- Which downstream output or activity is claimed to have used a recalled representation? relationship
- Who asserted use and what observation or receipt supports that assertion? evidence
- Does the evidence show inclusion, citation, assessed support, or only unknown use? classification
- How are absent, disputed or retracted use assertions represented without rewriting retrieval history? lifecycle
Trace governance Integrity, review, disclosure and lawful disposition.
Trace integrity and review
Assess completeness and consistency of this event record using recorded evidence. Record checks do not certify retriever quality, establish factual truth or reproduce unavailable runs. Safety issues are routed by reference at policy level.
Trace integrity and review
Assess completeness and consistency of this event record using recorded evidence. Record checks do not certify retriever quality, establish factual truth or reproduce unavailable runs. Safety issues are routed by reference at policy level.
- Which recorder, source receipts and integrity checks support this trace revision? provenance
- Which checks found unresolved identifiers, inconsistent ranks or unsupported use links? validation
- Which external assessment or incident record addresses disputed retrieval behavior? security
- Which record corrections supersede earlier statements and who authorized them? lifecycle
Trace disclosure and disposition
Queries, candidate identifiers, counts and use assertions may reveal sensitive material even without payload. Govern this trace with field-level views and external retention policy. Durable references do not require indefinite personal-data retention.
Trace disclosure and disposition
Queries, candidate identifiers, counts and use assertions may reveal sensitive material even without payload. Govern this trace with field-level views and external retention policy. Durable references do not require indefinite personal-data retention.
- Which roles may inspect each part of the trace and for what purpose? access
- Which query, score, candidate or use fields require redaction in each view? privacy
- What retention schedule, hold and disposition authority apply to this event record? retention
- How will authorized erasure preserve only permitted tombstones and avoid resurrecting source data? constraint
Classifiers Filled
- Family
- World Models
- Category
- Information and virtual systems
- Entry kind
- event
- Navigation path
- NAV.INF.MEM.RCL
- Domain
- INF.MEM.RCL
- Industry
- Cross-industry
- Tags
- memoryrecalleventinf.mem.rcl
What it is Filled
Format-independent event record for a bounded recall attempt, including failed, denied, empty or partial outcomes. It captures observations and references to external decisions rather than performing retrieval or certifying answer grounding.
In scope
- Occurrence identity, request intent and query transformation references.
- Reported authority, searched state, retrieval stages and completion qualifiers.
- Returned item/passage representations, ranks, score meanings and candidate lineage.
- Selection, handoff and attributed downstream-use evidence, including explicit unknowns.
- Governance, correction, disclosure and disposition of the recall trace itself.
Out of scope
- Biological memory, psychological recall or claims about a model internal cognitive state.
- Memory item formation, consolidation, trust judgment, conflict precedence and memory-store lifecycle.
- Index construction, embedding generation, retrieval/reranking execution and context packing.
- Answer generation, statement entailment, factual truth and causal attribution of private model reliance.
- Grant creation, policy enforcement, generic audit infrastructure and incident response execution.
Why it exists Filled
Describe one agent-memory retrieval attempt: its query, authorized scope, observed candidates and ranking, and evidence of downstream selection or use.
Distinguishing features Derived, awaiting review
- Unlike WM-MEM-001: Memory Store remains the master for tenancy, storage and store lifecycle; this event references the searched store and observed scope.
- Unlike WM-MEM-002: Memory Item remains the durable record master. Registry parent_ids is a candidate navigation relation, not event inheritance or mandatory containment. Empty and denied attempts can reference no item.
- Unlike WM-MEM-006: Memory Formation Event owns creation lineage; recall reads prior records and does not form or revise them.
- Unlike WM-MEM-009: Memory Conflict and Precedence Rule owns conflict resolution policy. The recall trace records external decision references without resolving contradictions.
- Unlike WM-MEM-010: Memory Retention and Forgetting Policy owns schedule and erasure rules; this event records its own disposition binding and executor receipt.
- Unlike WM-MEM-011: Memory Scope and Sharing Grant owns access authority. Recall records reported decisions and never creates grants.
- Unlike WM-MEM-013: Context Assembly owns context packing and actual inclusion. A selected or delivered candidate is not automatically included.
- Unlike WM-MEM-014: Retrieval Index owns indexing, chunking, embeddings and refresh lifecycle. Recall references configuration and observed snapshot only.
- Unlike WM-MEM-015: Retrieval Passage owns passage identity and source lineage. Recall carries a versioned passage locator and candidate role.
- Unlike WM-MEM-016: Answer Grounding owns claim-to-evidence support assessment. Recall links use receipts and grounding records without inferring entailment or causal model reliance.
- Unlike WM-MEM-017: Context Policy owns overriding context rules. Recall records applicable version and external decision receipt.
- Unlike WM-MEM-018: Memory Trust Assessment owns trust judgments; retrieval rank and score never substitute for trust.
- Unlike WM-MEM-019: Memory Poisoning Incident owns incident handling; this record provides restricted evidence links only.
- Unlike WM-MEM-020: Memory Quarantine and Redaction owns quarantine decisions; the recall trace records referenced exclusion or redaction receipts.
Note: Derived from boundary notes against neighbouring models.
What robots and AI may and may not do Derived, awaiting review
Must not
- Resolve versioned references under the recipient scope; return redacted or unresolved markers without leaking denied existence.
- Allow denied, empty and unobserved-result attempts; do not fabricate candidates, scores, times or answer use.
- Deny by default; require caller, tenant, purpose and current policy for every trace view and mutation.
May
- Register recall occurrence: Proposed local record operation, not implemented. Record a new observed attempt or recognize duplicate delivery.
- Attach execution observations: Proposed local record operation, not implemented. Add externally observed stages and completion evidence.
- Attach candidate and rank evidence: Proposed local record operation, not implemented. Bind returned representations and stage-specific ordering.
- Link selection and use receipts: Proposed local record operation, not implemented. Record attributed downstream evidence.
- Review trace consistency: Proposed local record operation, not implemented. Check this record, not runtime retrieval quality.
- Record correction or disposition receipt: Proposed local record operation, not implemented. Link an authorized correction, redaction or erasure outcome.
Note: Derived from functions, policies, CRUD and access rules; prohibitions were not authored for agents as such.
Moral aspects Derived, awaiting review
- Use a jurisdiction-appropriate retention and privacy profile before operational adoption; no universal duration or legal-compliance assertion.
- Adopters must review privacy, records, licensing, retention and cross-border rules for their own profile.
Note: Sentences mentioning harm, privacy, consent or similar, collected from the specification.
Owners Filled
Steward
Identify the accountable memory-system operator and trace custodian by role, with namespace and master-system binding.
Roles
- Trace custodian
- Own the trace namespace, record policy and disposition coordination.
- Recall recorder
- Attribute observed requests, stage receipts and unknowns without inventing outcomes.
- Authority reviewer
- Verify scope and referenced decisions; cannot create a grant through a trace patch.
- Evidence reviewer
- Inspect consistency, score interpretation and downstream-use evidence within granted scope.
- Disposition executor
- Apply authorized trace retention or erasure policy and return scoped completion receipts.
Links to other meta-models Filled
references
- WM-MEM-001 - Proposed optional neighbor binding pending version pin: Memory Store remains the master for tenancy, storage and store lifecycle; this event references the searched store and observed scope.
- WM-MEM-002 - Proposed optional neighbor binding pending version pin: Memory Item remains the durable record master. Registry parent_ids is a candidate navigation relation, not event inheritance or mandatory containment. Empty and denied attempts can reference no item.
- WM-MEM-006 - Proposed optional neighbor binding pending version pin: Memory Formation Event owns creation lineage; recall reads prior records and does not form or revise them.
- WM-MEM-009 - Proposed optional neighbor binding pending version pin: Memory Conflict and Precedence Rule owns conflict resolution policy. The recall trace records external decision references without resolving contradictions.
- WM-MEM-010 - Proposed optional neighbor binding pending version pin: Memory Retention and Forgetting Policy owns schedule and erasure rules; this event records its own disposition binding and executor receipt.
- WM-MEM-011 - Proposed optional neighbor binding pending version pin: Memory Scope and Sharing Grant owns access authority. Recall records reported decisions and never creates grants.
- WM-MEM-013 - Proposed optional neighbor binding pending version pin: Context Assembly owns context packing and actual inclusion. A selected or delivered candidate is not automatically included.
- WM-MEM-014 - Proposed optional neighbor binding pending version pin: Retrieval Index owns indexing, chunking, embeddings and refresh lifecycle. Recall references configuration and observed snapshot only.
- WM-MEM-015 - Proposed optional neighbor binding pending version pin: Retrieval Passage owns passage identity and source lineage. Recall carries a versioned passage locator and candidate role.
- WM-MEM-016 - Proposed optional neighbor binding pending version pin: Answer Grounding owns claim-to-evidence support assessment. Recall links use receipts and grounding records without inferring entailment or causal model reliance.
- WM-MEM-017 - Proposed optional neighbor binding pending version pin: Context Policy owns overriding context rules. Recall records applicable version and external decision receipt.
- WM-MEM-018 - Proposed optional neighbor binding pending version pin: Memory Trust Assessment owns trust judgments; retrieval rank and score never substitute for trust.
- WM-MEM-019 - Proposed optional neighbor binding pending version pin: Memory Poisoning Incident owns incident handling; this record provides restricted evidence links only.
- WM-MEM-020 - Proposed optional neighbor binding pending version pin: Memory Quarantine and Redaction owns quarantine decisions; the recall trace records referenced exclusion or redaction receipts.
aligned
- W3C PROV-DM - Optional conceptual mapping of attempt to activity, record to entity and evidenced usage; mapping tests remain pending.
- CloudEvents 1.0.2 - Optional transport envelope. Keep source/id delivery deduplication separate from attempt, retry and record revision identities.
- W3C Web Annotation Data Model - Optional passage selector/state mapping; preserve representation and timestamp profile differences.
neighbor
- WM-MEM-001 - Memory Store remains the master for tenancy, storage and store lifecycle; this event references the searched store and observed scope.
- WM-MEM-002 - Memory Item remains the durable record master. Registry parent_ids is a candidate navigation relation, not event inheritance or mandatory containment. Empty and denied attempts can reference no item.
- WM-MEM-006 - Memory Formation Event owns creation lineage; recall reads prior records and does not form or revise them.
- WM-MEM-009 - Memory Conflict and Precedence Rule owns conflict resolution policy. The recall trace records external decision references without resolving contradictions.
- WM-MEM-010 - Memory Retention and Forgetting Policy owns schedule and erasure rules; this event records its own disposition binding and executor receipt.
- WM-MEM-011 - Memory Scope and Sharing Grant owns access authority. Recall records reported decisions and never creates grants.
- WM-MEM-013 - Context Assembly owns context packing and actual inclusion. A selected or delivered candidate is not automatically included.
- WM-MEM-014 - Retrieval Index owns indexing, chunking, embeddings and refresh lifecycle. Recall references configuration and observed snapshot only.
- WM-MEM-015 - Retrieval Passage owns passage identity and source lineage. Recall carries a versioned passage locator and candidate role.
- WM-MEM-016 - Answer Grounding owns claim-to-evidence support assessment. Recall links use receipts and grounding records without inferring entailment or causal model reliance.
- WM-MEM-017 - Context Policy owns overriding context rules. Recall records applicable version and external decision receipt.
- WM-MEM-018 - Memory Trust Assessment owns trust judgments; retrieval rank and score never substitute for trust.
- WM-MEM-019 - Memory Poisoning Incident owns incident handling; this record provides restricted evidence links only.
- WM-MEM-020 - Memory Quarantine and Redaction owns quarantine decisions; the recall trace records referenced exclusion or redaction receipts.
parent
- WM-MEM-002
What else AI and robots need to interact with it Incomplete
Identity and identifiers required Filled
- Authoritative master-system identifier
- Governed global identifier or IRI
- Dimension-assigned UUID or ULID
Direct properties not applicable Not applicable
Not applicable
Institutional or informational subject: no invented physical properties.
Recognition optional Missing, in the backlog
Not described yet. This gap is in the card backlog.
Capabilities and actions required Filled
- Register recall occurrence: Proposed local record operation, not implemented. Record a new observed attempt or recognize duplicate delivery.
- Attach execution observations: Proposed local record operation, not implemented. Add externally observed stages and completion evidence.
- Attach candidate and rank evidence: Proposed local record operation, not implemented. Bind returned representations and stage-specific ordering.
- Link selection and use receipts: Proposed local record operation, not implemented. Record attributed downstream evidence.
- Review trace consistency: Proposed local record operation, not implemented. Check this record, not runtime retrieval quality.
- Record correction or disposition receipt: Proposed local record operation, not implemented. Link an authorized correction, redaction or erasure outcome.
Hazards and failure modes optional Missing, in the backlog
Not described yet. This gap is in the card backlog.
Standards and interfaces required Derived, awaiting review
- PROV-DM: The PROV Data Model
- RFC 3339: Date and Time on the Internet: Timestamps
Context of use required Filled
- No jurisdiction or sector is assumed. Adopters must review privacy, records, licensing, retention and cross-border rules for their own profile.
- Open specifications are conceptual alignments; Apache API examples are not mandatory runtime choices.
Sources Filled
- PROV-DM: The PROV Data Model - W3C
- CloudEvents Specification - CNCF
- Apache Lucene TopDocs - Apache Software Foundation
- Apache Lucene TotalHits - Apache Software Foundation
- Apache Solr Common Query Parameters - Apache Software Foundation
- Web Annotation Data Model - W3C
- RFC 3339: Date and Time on the Internet: Timestamps - IETF
- OWASP LLM08:2025 Vector and Embedding Weaknesses - OWASP
Open questions
- Run coordinator source checks and independently verify cited claims, current applicable versions, immutable documentation pins and errata.
- Develop pinned neighbor contracts, nested instance schemas and fixtures for duplicate delivery, retry, denial, timeout, lower-bound counts, rank fusion, revoked cache access, absent use evidence and lawful erasure.
- Test PROV-DM, CloudEvents and Web Annotation projections with explicit identity and timestamp rules and declared projection losses.
- Restore independent external review before any canonical or publishable-draft promotion.
- Nested candidate, stage, score and use-assertion schemas with executable semantic constraints are pending.
- Pinned sibling specifications, transport mappings, replay fixtures and multimodal locator profiles are pending.
- No independent second-provider review; a separate local no-tools self-audit is required.
- Direct HTTP checks are unattempted under the owner-reported sandbox restriction; web-service readings do not establish origin HTTP status or current version completeness.
- No benchmark establishes retrieval quality, factual accuracy or causal answer reliance.
Machine files
Provenance
world-models research · reviewable-draft
Built from: models/wm-mem-007-memory-recall-event/spec.yaml