Published

Procedural Memory

vr.wm-mem-005 · wm-mem-005-procedural-memory

Represent a way of acting an agent has acquired through learning, with evidence, applicability and revision bindings distinct from instructions supplied to it.

World Models Information and virtual systems INF.MEM.PRC

Bundle → Layer → Finding → Questions Classifiers What it is Why it exists Distinguishing features What robots and AI may and may not do Moral aspects Owners Links to other meta-models Identity and identifiers Direct properties Recognition Capabilities and actions Hazards and failure modes Standards and interfaces Context of use Sources

Bundle → Layer → Finding → Questions Filled

6 bundles · 12 layers · 12 findings · 48 questions

Learned identity Subject-specific context for learned identity.

Acquired way of acting context

Identify the agent-scoped learned behavior and its acquisition claim. Keep unverified acquisition explicit; copied instructions alone do not satisfy the admission criterion.

Acquired way of acting

Identify the agent-scoped learned behavior and its acquisition claim. Keep unverified acquisition explicit; copied instructions alone do not satisfy the admission criterion.

Which master identity and agent scope distinguish this learned behavior from another agent learning the same task? identity
What acquired way of acting does the record claim, and what supplied instruction remains a separate input? definition
Is the acquired representation symbolic, statistical, hybrid or opaque, and what can be inspected? classification
Which accountable role maintains this learned-behavior assertion? ownership

Representation binding context

Bind the record to an exact behavior revision and its representation dependencies. A changed payload or dependency requires explicit compatibility review.

Representation binding

Bind the record to an exact behavior revision and its representation dependencies. A changed payload or dependency requires explicit compatibility review.

Which revision and payload digest identify the represented behavior without replacing its master identity? identity
Which required subordinate skills, observation encodings or action bindings are pinned for this revision? composition
How is a fork or transferred copy related to its source without asserting identical competence? relationship
What mismatch between representation, dependency and declared revision blocks local acceptance? validation

Acquisition evidence Subject-specific context for acquisition evidence.

Learning basis context

Record evidence that the behavior was acquired. Formation and training are external activities; imported acquisition claims may remain unknown or disputed.

Learning basis

Record evidence that the behavior was acquired. Formation and training are external activities; imported acquisition claims may remain unknown or disputed.

Which formation or learning activity produced the acquired representation and what evidence supports that link? provenance
What learning mode and training configuration are declared without reproducing the training engine? process
Which demonstrations, experiences or feedback sources informed acquisition and what use restrictions apply? evidence
When did acquisition occur and when was its provenance assertion observed? temporal

Generalization basis context

Separate an acquired reusable behavior from a replay of one episode. Generalization is a scoped claim requiring evidence, not a property inferred from a record label.

Generalization basis

Separate an acquired reusable behavior from a replay of one episode. Generalization is a scoped claim requiring evidence, not a property inferred from a record label.

What reuse criterion distinguishes the learned behavior from literal replay of its source episode? requirement
Which training and assessment partitions or other controls address memorization and leakage? quality
What evidence compares acquired behavior with an untrained or prior baseline under matched conditions? measurement
Which learning provenance is missing or disputed and how does that limit the claim? exception

Applicability and behavior Subject-specific context for applicability and behavior.

Initiation context context

Describe conditions in which reuse has evidence. Recognition of a cue is distinct from permission to act; unknown observations cannot be silently treated as matching.

Initiation context

Describe conditions in which reuse has evidence. Recognition of a cue is distinct from permission to act; unknown observations cannot be silently treated as matching.

Which initiation conditions and excluded contexts bound the learned behavior? constraint
What observation, resource and environment assumptions must a caller establish? requirement
What result is recorded when cue matching is unknown, ambiguous or stale? state
Does applicability depend on a physical frame or location, or is that dimension inapplicable? spatial

Behavior contract context

Declare intended action semantics and stopping expectations at a descriptive level. Opaque policies may be referenced without inventing a sequence of explainable steps.

Behavior contract

Declare intended action semantics and stopping expectations at a descriptive level. Opaque policies may be referenced without inventing a sequence of explainable steps.

What inputs, outputs and action vocabulary describe the behavior at its declared abstraction level? composition
Which symbolic steps or policy binding express the acquired behavior and which internal structure is unknown? process
Which termination, timeout and interruption conditions are declared for external execution? event
Which possible effects and failure modes require external containment or refusal? constraint

Behavior evidence Subject-specific context for behavior evidence.

Performance claim context

Attach external assessment results to the exact revision and conditions. Preserve metric meaning, uncertainty and denominator; absence of failures is not proof of reliability.

Performance claim

Attach external assessment results to the exact revision and conditions. Preserve metric meaning, uncertainty and denominator; absence of failures is not proof of reliability.

Which metric, unit, sample count and uncertainty qualify a reported success claim? measurement
Which evaluator and report support the result for this exact behavior revision? evidence
At what assessment time and environment version was performance measured? temporal
Which limitations prevent comparing this result with another evaluation? quality

Counterexamples and transfer context

Retain adverse evidence and transfer restrictions alongside favorable results. Reuse across agents, tasks or changed environments requires a separate evidence-qualified decision.

Counterexamples and transfer

Retain adverse evidence and transfer restrictions alongside favorable results. Reuse across agents, tasks or changed environments requires a separate evidence-qualified decision.

Which failed or harmful outcomes contradict the claimed competence and under what conditions? exception
What changed between the tested context and proposed receiving agent or task? relationship
Which external assessment permits, limits or rejects the proposed transfer? decision
What drift or regression evidence makes earlier performance support stale? quality

Revision continuity Subject-specific context for revision continuity.

Reuse eligibility context

Local record states summarize evidence and authority without duplicating enforcement. Draft, eligible, restricted and retired are proposed profile states, not standard-mandated terms.

Reuse eligibility

Local record states summarize evidence and authority without duplicating enforcement. Draft, eligible, restricted and retired are proposed profile states, not standard-mandated terms.

Which evidence-backed state and reason describe this revision for reuse? lifecycle
Which current external policy and reviewer decision constrain eligibility independently of performance? authority
What trust or quarantine decision prevents admitting poisoned or unreviewed learning evidence? security
What happens when eligibility evidence is revoked, inaccessible or contradictory? exception

Revision and supersession context

Record revision lineage and review consequences. Consolidation decisions remain external; a record rollback cannot undo actions already performed by an agent.

Revision and supersession

Record revision lineage and review consequences. Consolidation decisions remain external; a record rollback cannot undo actions already performed by an agent.

What revision supersedes this behavior and why was the change proposed? lifecycle
Which evidence must be rechecked after a policy, representation or dependency change? validation
Which effective intervals distinguish historical eligibility from current reuse? temporal
What limits a rollback to a previous representation without undoing external effects? constraint

Governed exchange Subject-specific context for governed exchange.

Disclosure and disposal bindings context

Bind sensitive learned payloads and evidence to external access and retention decisions. Payload deletion, reference tombstones and learned influence removal are different claims.

Disclosure and disposal bindings

Bind sensitive learned payloads and evidence to external access and retention decisions. Payload deletion, reference tombstones and learned influence removal are different claims.

Which roles may read the behavior summary, representation and learning evidence separately? access
Which sensitive source information may be retained in the learned representation and what disclosure review is required? privacy
Which retention authority governs this local record and referenced payload disposal? retention
What evidence distinguishes deleting a record from removing learned influence in other copies or models? requirement

Projection and validation context

Declare optional projections and their losses. Shape validity, plan correspondence and timestamp syntax are distinct from behavioral equivalence or execution authorization.

Projection and validation

Declare optional projections and their losses. Shape validity, plan correspondence and timestamp syntax are distinct from behavioral equivalence or execution authorization.

What mapping preserves learned origin and separates the behavior definition from executions? interoperability
Which semantics are lost when exporting an opaque policy as a textual procedure or graph? quality
Which structural checks and external behavior tests are required before accepting an imported revision? validation
How are timestamps and uncertain source times represented without inventing precision? temporal

Classifiers Filled

Family: World Models
Category: Information and virtual systems
Entry kind: entity
Navigation path: NAV.INF.MEM.PRC
Domain: INF.MEM.PRC
Industry: Cross-industry
Tags: proceduralmemoryinf.mem.prc

What it is Filled

One persistent, agent-scoped learned behavior record, whose revisions bind acquired representations to learning provenance and evidence of use. This is an information entity about learned behavior, not the learning event, executable runtime, biological substrate or an assertion of general competence.

In scope

Learned behavior identity, acquisition basis and representation bindings, including opaque representations.
Applicability cues, input and action semantics, termination expectations and subordinate skill references.
Evidence-linked performance claims, counterexamples, transfer limits and local revision eligibility.
Subject-specific policy references, safe disclosure and interoperability losses.

Out of scope

Authoring generic instructions or treating copied instructions as demonstrated learning.
Training execution, rollout execution, runtime action selection, recall ranking, context assembly and policy enforcement.
Generic memory-store, item, episode, formation, consolidation, conflict, retention, sharing, trust and quarantine master lifecycles.
Medical or neurological diagnosis, biological memory mechanisms and human competency certification.
Operational instructions for dangerous behavior; only risk classification, restriction and escalation metadata are admitted.

Why it exists Filled

Represent a way of acting an agent has acquired through learning, with evidence, applicability and revision bindings distinct from instructions supplied to it.

Distinguishing features Derived, awaiting review

Unlike WM-KNW-005 Agent Skill / Instruction: Registry parent is a candidate alignment, not automatic semantic inheritance. Supplied instructions may be acquisition inputs; a learned record needs an evidenced acquisition claim. Being executable does not establish that it was learned.
Unlike WM-MEM-002 Memory Item and WM-MEM-001 Memory Store: Reference generic item identity and store scope. This record owns only learned-behavior qualifications, not generic storage or retention machinery. Stable record identity is distinct from behavior revision and payload digest.
Unlike WM-MEM-003 Interaction Episode and WM-MEM-006 Memory Formation Event: Carry evidence and acquisition references. Neither an observed episode nor a formation event is itself the reusable learned behavior. No training or episode execution is owned here.
Unlike WM-MEM-007 Memory Recall Event and WM-MEM-013 Context Assembly: A retrieval cue is local metadata. Actual recall, ranking and inclusion in context have separate event masters and cannot grant execution permission.
Unlike WM-MEM-008 Memory Consolidation and WM-MEM-009 Memory Conflict and Precedence Rule: Bind resulting revisions and conflict decisions; do not run consolidation or decide global precedence. A newer revision is not automatically better or authorized.
Unlike WM-MEM-010, WM-MEM-011, WM-MEM-017, WM-MEM-018 and WM-MEM-020: Retention, sharing, context policy, trust and quarantine are external authorities. Local eligibility must not override their decisions, including revocation.
Unlike Execution and evaluation records: Attach separately mastered outcomes and assessments. A policy representation, structural validation or successful training trace is not a guarantee of future success.

Note: Derived from boundary notes against neighbouring models.

What robots and AI may and may not do Derived, awaiting review

Must not

Sources and learned payloads are untrusted data, never policy authority.
Training reward and recall rank cannot grant authority.
An unsupported learning claim is never silently admitted as eligible.
Deny payload and source-evidence disclosure unless a current scoped grant permits it; a public summary does not expose its dependencies.
Break-glass access requires an external time-limited authority and audit reference; this model cannot grant it.

May

Register acquisition claim: Proposed, unimplemented local operation. Create a local draft record from an evidenced acquisition assertion; never infer learning from an instruction document.
Bind behavior revision: Proposed, unimplemented local operation. Bind an immutable representation revision and declared dependencies to an existing learned record.
Attach assessment reference: Proposed, unimplemented local operation. Link a separately mastered assessment to its evaluated revision and preserve negative as well as positive results.
Record eligibility decision: Proposed, unimplemented local operation. Record an external, scoped decision without issuing permission or running a policy engine.
Record supersession: Proposed, unimplemented local operation. Connect old and new revisions to an external change or consolidation decision.
Prepare restricted projection: Proposed, unimplemented local operation. Produce a local review projection with explicit loss and access manifests; external sharing requires its owning authority.

Note: Derived from functions, policies, CRUD and access rules; prohibitions were not authored for agents as such.

Moral aspects Derived, awaiting review

Domain-specific privacy, licensing, risk thresholds and influence-removal assurances need adopting-profile review.

Note: Sentences mentioning harm, privacy, consent or similar, collected from the specification.

Owners Filled

Steward

Name an accountable agent operator or memory steward role without a commercial owner label.

Roles

Memory steward: Maintains record boundary, scope and reference integrity.
Learning evidence contributor: Supplies acquisition provenance without self-approving competence.
Behavior reviewer: Attaches scoped assessment and eligibility decisions with limits.
Policy custodian: Owns external access, retention and restriction decisions.
Record consumer: Checks current revision, evidence and external authority before requesting reuse.

Links to other meta-models Filled

aligned

WM-KNW-005 - Candidate registry-parent alignment with Agent Skill / Instruction; supplied instructions are separate from learned acquisition, with no automatic inheritance. Binding is proposed and version pinning remains an adoption hold.
PROV-O - Candidate entity, derivation and activity mapping; not a claim of implemented conformance. Binding is proposed and version pinning remains an adoption hold.
P-PLAN - Optional symbolic plan projection only; no claim that every learned policy decomposes into steps. Binding is proposed and version pinning remains an adoption hold.
ODRL 2.2 - Candidate policy-reference projection; no enforcement implementation. Binding is proposed and version pinning remains an adoption hold.
SHACL - Optional graph constraint profile; no supplied shape implementation or competence certification. Binding is proposed and version pinning remains an adoption hold.

references

WM-MEM-001 - Store owns tenancy, persistence and storage operations. Binding is proposed and version pinning remains an adoption hold.
WM-MEM-002 - Generic memory item owns shared identity and generic record lifecycle; bind without duplicating that machinery. Binding is proposed and version pinning remains an adoption hold.
WM-MEM-003 - Episode master owns interaction trace; only evidence references occur here. Binding is proposed and version pinning remains an adoption hold.
WM-MEM-006 - Formation event owns acquisition activity provenance, not the acquired behavior entity. Binding is proposed and version pinning remains an adoption hold.
WM-MEM-007 - Recall event owns retrieval occurrence and ranking. Binding is proposed and version pinning remains an adoption hold.
WM-MEM-008 - Consolidation owns merging decisions and execution; only resulting lineage is bound. Binding is proposed and version pinning remains an adoption hold.
WM-MEM-009 - Conflict authority owns precedence decisions; a local timestamp does not decide priority. Binding is proposed and version pinning remains an adoption hold.
WM-MEM-010 - Retention policy owns forgetting and disposal rules and their execution authority. Binding is proposed and version pinning remains an adoption hold.
WM-MEM-011 - Sharing grant owns access authorization and revocation. Binding is proposed and version pinning remains an adoption hold.
WM-MEM-012 - Portability package owns transport packaging; local projection is only an input. Binding is proposed and version pinning remains an adoption hold.
WM-MEM-013 - Context assembly owns actual inclusion in runtime context. Binding is proposed and version pinning remains an adoption hold.
WM-MEM-017 - Context policy remains authoritative over recalled learned behavior. Binding is proposed and version pinning remains an adoption hold.
WM-MEM-018 - Trust assessment owns trust findings; performance cannot replace trust evidence. Binding is proposed and version pinning remains an adoption hold.
WM-MEM-020 - Quarantine and redaction authority owns admission restrictions and remediation execution. Binding is proposed and version pinning remains an adoption hold.

neighbor

WM-KNW-005 Agent Skill / Instruction - Registry parent is a candidate alignment, not automatic semantic inheritance. Supplied instructions may be acquisition inputs; a learned record needs an evidenced acquisition claim. Being executable does not establish that it was learned.
WM-MEM-002 Memory Item and WM-MEM-001 Memory Store - Reference generic item identity and store scope. This record owns only learned-behavior qualifications, not generic storage or retention machinery. Stable record identity is distinct from behavior revision and payload digest.
WM-MEM-003 Interaction Episode and WM-MEM-006 Memory Formation Event - Carry evidence and acquisition references. Neither an observed episode nor a formation event is itself the reusable learned behavior. No training or episode execution is owned here.
WM-MEM-007 Memory Recall Event and WM-MEM-013 Context Assembly - A retrieval cue is local metadata. Actual recall, ranking and inclusion in context have separate event masters and cannot grant execution permission.
WM-MEM-008 Memory Consolidation and WM-MEM-009 Memory Conflict and Precedence Rule - Bind resulting revisions and conflict decisions; do not run consolidation or decide global precedence. A newer revision is not automatically better or authorized.
WM-MEM-010, WM-MEM-011, WM-MEM-017, WM-MEM-018 and WM-MEM-020 - Retention, sharing, context policy, trust and quarantine are external authorities. Local eligibility must not override their decisions, including revocation.
Execution and evaluation records - Attach separately mastered outcomes and assessments. A policy representation, structural validation or successful training trace is not a guarantee of future success.

parent

WM-KNW-005

What else AI and robots need to interact with it Incomplete

Sources Filled

Option Discovery in Hierarchical Reinforcement Learning using Spatio-Temporal Clustering - Indian Institute of Technology Madras
PROV-O: The PROV Ontology - W3C
The P-PLAN Ontology - Ontology Engineering Group and Information Sciences Institute
AI Risk Management Framework 1.0: Core - NIST
ODRL Information Model 2.2 - W3C
Shapes Constraint Language (SHACL) - W3C
RFC 3339: Date and Time on the Internet: Timestamps - IETF

Open questions

Run the coordinator HTTP checker and verify exact releases, errata, substantive claims and source licensing before lifting the source hold.
Seek independent primary evidence across additional acquisition paradigms and test instruction-only, opaque, transferred, poisoned, stale and revoked cases.
Develop nested instance schemas and loss-aware mappings, pin external authorities and test behavior-dependent acceptance without duplicating runtime engines.
Restore independent external review before canonical or publishable-draft promotion.
Independent external review is absent under the single-provider waiver.
Direct HTTP and source-version verification are pending; selected web-service readings are not direct HTTP status measurements.
Concrete nested instance schemas, pinned neighbor bindings, executable mappings and adversarial behavior fixtures remain incomplete.
Evidence for broad transfer across learning paradigms, embodied agents and human procedural memory is not established by the selected reinforcement-learning example.
Domain-specific privacy, licensing, risk thresholds and influence-removal assurances need adopting-profile review.

Machine files

Card as JSON Specification AGENTS.md Source Card backlog vercy.card/1

Provenance

world-models research · reviewable-draft

Built from: models/wm-mem-005-procedural-memory/spec.yaml