03_DAMS_V5_Spec_Outline.md
Memory Rebuild Docs/Memory Rebuild Review Packs/Archived Memory Rebuild Zips/DOC80_Memory_Control_Plane_PreSpec_Review_Pack_v1_0_2026-05-25/03_DAMS_V5_Spec_Outline.md
ELNOR REPO READER TEXT MIRROR Original path: Memory Rebuild Docs/Memory Rebuild Review Packs/Archived Memory Rebuild Zips/DOC80_Memory_Control_Plane_PreSpec_Review_Pack_v1_0_2026-05-25/03_DAMS_V5_Spec_Outline.md Source repo: /Users/OpenClaw1/Elnor/Elnor Specs Git branch: main Git commit: dbaa25962edc11ab30e8d4ca1715f9ae5bf77331 Generated: 2026-06-09T01:23:58.539Z --- # DAMS V5 / Memory Control Plane — Proposed Specification Outline **Document type:** Outline for the real spec. **Status:** Review draft. **Purpose:** Provide a spec-ready structure for review before drafting the full DAMS V5 / Memory Control Plane specification. --- ## Proposed title ```text DAMS V5 / Memory Control Plane Specification ``` Subtitle: ```text Canonical Memory Objects, Extraction, Scope Coordination, Context Delivery, and DAMS Attenuator Substrate ``` --- ## 0. Status, scope, and owner boundaries ### 0.1 Document status - Real specification draft target. - Supersedes DAMS V4.1 proposal on adoption. - DAMS retained as lineage name only. - DAMS is demoted to a substrate inside the Memory Control Plane. ### 0.2 Non-goals The spec does not: - replace DOC72 durable graph ownership; - replace EC as sole durable writer; - replace EC/PropA policy evaluation; - replace DOC24 packet assembly; - replace KDA rendering; - replace DOC11/OpenClaw runtime truth; - replace DOC23 task execution; - make Project mode mandatory for good memory. ### 0.3 Owner boundaries Define consumption and ownership map: ```text DOC72 — knowledge shape and graph payload contracts. DOC73 — deep source/corpus extraction, CU/source-bound synthesis, artifact/corpus surfaces. DOC25 — document ingestion, source artifacts, artifact segments, materialization. EC — sole durable writer, effective runtime truth, policy compilation/evaluation, command closure. PropA — semantic policy inputs, source rules, visibility, sensitivity, collection policy. DOC24 — packet lifecycle, delivery, injection manifests, delivery vocabulary. KDA — deterministic rendering from prepared bundles, manifest patches. DOC11/OpenClaw — final runtime prompt truth. BDSM/DOC8 — utility learning, attribution computation, false-suppression/replay. DOC23 — task execution; task outputs may feed memory but tasks are not memory organization. DOC20/Q — user-facing surfaces and inspectors. ``` ### 0.4 Terms retired or not introduced ```text Focus — not a new primitive. Work Scope — not a new user-visible primitive. Matter — legal-domain alias, not canonical domain-neutral term. Corpus — internal/source-processing term behind user-visible Library. PremiseFamily — legacy alias if retained; canonical V5 term should be Assertion. Claim — namespaced only. ``` --- ## 1. Governing invariants ### 1.1 No phantom seams Every control, schema, route, prompt product, and learning signal must map to a real owner and execution path. ### 1.2 EC sole durable writer All durable writes route through EC commands. ### 1.3 Policy-before-flow No extraction, rendering, export, delegation, carryover, or learning action proceeds without a policy decision where required. ### 1.4 Final-prompt proof before attribution No delivery utility or injection-learning signal flows from content absent from final-prompt delivery proof. ### 1.5 Role/action separation Object role and use warrant are separate. ### 1.6 Extraction route does not define canonical truth Same substantive assertion resolves to the same canonical Assertion pipeline regardless of extraction route. ### 1.7 Project optionality Project mode improves scoping and capture but is not required for high-quality memory. ### 1.8 Hot-path determinism No hot-path LLM calls, unbounded scans, or hidden network calls for runtime injection assembly. --- ## 2. Functional planes ### 2.1 Source and Evidence Plane Defines: - SourceEnvelope - SourceAuthorityClass - SourceSegment / SourceSpan references - Source policy tags - MaterializationState - PromptInjectionRiskFlags - EvidenceSupportEdge ### 2.2 Stored Memory Object Plane Defines: - MemoryObjectBase - MemoryObjectRole - Assertion / AssertionVariant - EvidenceRecord - ConsolidatedUnderstanding reference contract - Directive - Procedure - NullResultMemory - IssueFrame / WorkingStateEvent - WorkEpisode / RecentActivityRollup - IncidentObservation / FrictionEvent ### 2.3 Organization and User Interaction Plane Defines: - Project mode consumption boundary - Library - TopicLens - TopicCollectionDirective - SearchAffordance - Inspector read-models ### 2.4 Temporal and Working-State Plane Defines: - WorkSession - WorkEpisode - EpisodeSegment - EpisodePolicyEpoch - RecentActivityRollup consumption discipline - IssueFrame - WorkingStateEvent / IssueFrameUpdate - CognitiveDiff ### 2.5 Policy and Permission Plane Defines: - MemoryPolicyDecision - PolicyObligation - PolicyStamp - PolicyStampInvalidation - policy-stage 2×2 matrix - meet/composition by dimensions - UI privacy/control implications ### 2.6 Scope Resolution and Coordination Plane Defines: - ScopeIdentityRoot - ScopeEquivalenceBinding - ScopeContainerRelation - ScopeMembrane - ScopeAffinity - ScopeResolutionTrace - ScopePopulationHealth ### 2.7 Extraction and Ingestion Plane Defines: - ExtractionContextPlan - ExtractionRouteContext - ExtractionOutputKind - AssertionCandidateDisposition - extraction depth - collection mode - Topic-driven extraction - Library/corpus-driven extraction ### 2.8 Runtime Assembly and Delivery Plane Defines: - MemoryContextPlan - context products - Memory Use Contract - product prompt shells - DOC24/KDA/DOC15/DOC11 handoff - ContextPacketProof / MemoryFlowCertificate ### 2.9 Learning and Evaluation Plane Defines: - context-product learning targets - false-suppression sampling - injection product utility - organization/membership learning - extraction learning - UI learning - policy/classification learning - DOC8/BDSM boundary --- ## 3. Source and Evidence Plane ### 3.1 SourceEnvelope ### 3.2 SourceAuthorityClass ### 3.3 SourceSpan / ArtifactSegment support ### 3.4 EvidenceSupportEdge ### 3.5 Source materialization and quality ### 3.6 Source policy and prompt-injection risk ### 3.7 Source dedupe Rules: ```text Same source/span should dedupe by source hash, content hash, citation/authority ID, segment ID, and version. ``` --- ## 4. Canonical Assertion pipeline ### 4.1 Assertion definition ### 4.2 AssertionCandidate ### 4.3 AssertionResolution ### 4.4 AssertionVariant ### 4.5 AssertionTemporalProfile ### 4.6 AssertionLifecycleState ### 4.7 AssertionStalenessState ### 4.8 AssertionUseWarrant ### 4.9 Evidence support and contrary support ### 4.10 Dedupe / canonical proposition key ### 4.11 Promotion rules ### 4.12 Degradation rules ### 4.13 Assertion vs Claim / Premise / CU mapping Include normative table: | Existing term | V5 canonical treatment | |---|---| | Claim | namespaced only or AssertionCandidate | | Premise | use role for Assertion | | PremiseFamily | legacy alias for Assertion if needed | | DOC73 VersionedClaim | map to AssertionVariant, CU component, or evidence-only depending use | | ConsolidatedUnderstanding | source-bound synthesis | | KDA rendered claim | rendering only, not new truth | | DOC23 claim extractor output | evaluation unit unless promoted through Assertion pipeline | | BDSM learned pattern assertion | learning signal, not truth unless promoted | --- ## 5. ConsolidatedUnderstandings and source-bound synthesis ### 5.1 CU definition ### 5.2 CU source-span requirement ### 5.3 CU authority and freshness ### 5.4 CU-to-Assertion candidate generation ### 5.5 CU injection rules Normative rule: ```text CU may inject as source-bound synthesis. CU may not substitute for reusable Assertion unless Assertions have been resolved from it. ``` --- ## 6. Directives, procedures, and non-assertion memory ### 6.1 Directive Examples: ```text Call me Will. Prefer concise red-team cards. Do not use markdown for this current draft unless user saves durable preference. ``` ### 6.2 Procedure ### 6.3 Standing procedure ### 6.4 Directive/procedure extraction rules ### 6.5 Directive/procedure injection rules ### 6.6 Non-assertion fallback --- ## 7. Temporal and working-state memory ### 7.1 WorkSession ### 7.2 WorkEpisode ### 7.3 EpisodeSegment ### 7.4 EpisodePolicyEpoch ### 7.5 RecentActivityRollup consumption Rule: ```text RecentActivityRollup may orient query framing/resume but must not satisfy evidence or authority queries. ``` ### 7.6 IssueFrame ### 7.7 WorkingStateEvent / IssueFrameUpdate ### 7.8 CognitiveDiff ### 7.9 Resume / recent work context products --- ## 8. Project, Library, and Topic ### 8.1 Project mode - User-visible when active. - Per surface, not global. - No global focus mode. - Does not create a separate memory store. - Improves scoping, capture, Library binding, and injection coordination. - Not required for memory quality. ### 8.2 Library - User-visible source collection. - May be linked, indexed, deeply ingested, extracted, stale, selected, included. - Internal corpus/read-model states are hidden unless inspector. ### 8.3 TopicLens - User-visible semantic lens. - Owns no raw truth. - References Assertions, CUs, EvidenceRecords, Directives, Procedures, IssueFrames, sources, etc. - Supports Topic Notice and Topic Slice. ### 8.4 TopicCollectionDirective - Separate governed extraction rule. - May drive backfill, selected-source extraction, or future watch. - Must be policy-gated. - Emits AssertionCandidates / EvidenceRecords / memberships. ### 8.5 Topic vs Library vs CU vs Assertion examples ### 8.6 Topic coverage health --- ## 9. Policy and permission ### 9.1 Policy-stage 2×2 matrix Dimensions: ```text intrinsic vs session write-stage vs retrieval/render-stage ``` ### 9.2 MemoryPolicyDecision ### 9.3 PolicyObligation ### 9.4 Policy meet / composition ### 9.5 PolicyStamp ### 9.6 PolicyStampInvalidation ### 9.7 UI/privacy settings and effective state ### 9.8 Carryover/export/delegation policy ### 9.9 Learning partition policy --- ## 10. Scope resolution and coordination ### 10.1 ScopeIdentityRoot ### 10.2 ScopeEquivalenceBinding ### 10.3 ScopeContainerRelation ### 10.4 ScopeMembrane ### 10.5 ScopeAffinity ### 10.6 ScopePopulationHealth ### 10.7 ScopeResolutionTrace ### 10.8 Capacity collapse and DAMS scoring ### 10.9 Cross-scope / ethical-wall rule ### 10.10 Scope in extraction, injection, UI, and learning --- ## 11. Extraction and ingestion ### 11.1 ExtractionContextPlan ### 11.2 ExtractionRouteContext ### 11.3 ExtractionOutputKind ### 11.4 AssertionCandidateDisposition ### 11.5 Extraction depth and collection mode ### 11.6 Library/corpus extraction ### 11.7 Topic-driven extraction ### 11.8 Project-mode capture ### 11.9 Chat capture ### 11.10 Task-output extraction ### 11.11 Incident/friction routing ### 11.12 Review queues ### 11.13 Source-to-memory receipts --- ## 12. Runtime assembly and delivery ### 12.1 MemoryContextPlan ### 12.2 Context product registry Products: ```text Direct Memory Item Assertion Packet Topic Notice Topic Slice Library Notice Library Source Slice Recent Work Orientation IssueFrame Orientation Directive Block Procedure Block Warning / Constraint Search Affordance Blocked / Reference-Only Notice ``` ### 12.3 Memory Use Contract ### 12.4 Product prompt shells ### 12.5 Placement and salience rules ### 12.6 Cap-and-trade / role-band budget allocation ### 12.7 Topic Notice vs Topic Slice decision rules ### 12.8 Library Notice vs Library Source Slice decision rules ### 12.9 Assertion Packet warrant rules ### 12.10 Recent Work / IssueFrame injection rules ### 12.11 DOC24 handoff ### 12.12 KDA render constraints ### 12.13 DOC11/OpenClaw final prompt truth ### 12.14 ContextPacketProof --- ## 13. DAMS attenuator and capacity-prior substrate ### 13.1 AttenuatorContributor ### 13.2 Channel tags ```text standing | learning | template | project | topic | library | source | directive | procedure | policy_warning ``` ### 13.3 Signed additive-deviation operator Use tanh soft-knee open-interval formulation. ### 13.4 Scope-equivalence collapse ### 13.5 Signal-family caps ### 13.6 Channel-scoped false-suppression floor ### 13.7 Role-band allocation ### 13.8 Suppression sampling ### 13.9 `ResolvedAttenuator` ### 13.10 Inspection fields --- ## 14. Memory proof and manifests ### 14.1 MemoryFlowCertificate ### 14.2 ContextPacketProof ### 14.3 CarryoverCapsule ### 14.4 SuppressionVisibility ### 14.5 Final-prompt proof and attribution gating ### 14.6 Certificate visibility and metadata leakage --- ## 15. Learning and evaluation ### 15.1 Learning target taxonomy ```text storage_extraction classification_policy organization_membership scope_resolution topic_notice_vs_slice library_notice_vs_source_slice packet_length prompt_shell warrant_assignment ui_surface ``` ### 15.2 DOC8 / Learning Engine boundary ### 15.3 BDSM utility boundary ### 15.4 Final prompt delivery proof ### 15.5 False-suppression sampling ### 15.6 Counterfactual replay design ### 15.7 Product-level learning ### 15.8 UI learning ### 15.9 Policy learning restrictions --- ## 16. UI and inspector surfaces ### 16.1 User-visible concepts Default visible concepts: ```text Project, when active Library, when selected/open/included Topic Notice, when useful Recent Work / Resume Card, when useful Search Inspector / Why included ``` Hidden unless inspector: ```text ScopeRoot / scope internals WorkEpisode internals IssueFrame internals Policy generation Manifests KDA patches MemoryFlowCertificate ``` ### 16.2 Why included ### 16.3 Why excluded ### 16.4 Why not remembered ### 16.5 Topic page ### 16.6 Library page ### 16.7 Privacy and policy banners ### 16.8 Search-more affordances --- ## 17. Hot-path cost budget ### 17.1 HotPathCostBudget schema ### 17.2 No hot-path LLM calls ### 17.3 No unbounded scans ### 17.4 Local fallback ### 17.5 Cache degradation behavior --- ## 18. Research-to-requirements matrix Include concise matrix: ```text RAG → provenance/source-backed products Lost-in-the-Middle → no long undifferentiated dumps; salient placement Self-RAG → adaptive retrieval / notice / ask / suppress, not fixed top-k MemGPT → tiered paging through Topic Notice, Library Notice, search affordances, carryover ``` --- ## 19. Acceptance fixtures Minimum fixtures: 1. Qwen not working — ephemeral incident. 2. Qwen locked model — rolling operational Assertion. 3. Memory system broken because specs were underspecified — IssueFrame + candidate Assertion. 4. Ninth Circuit scienter — same Assertion from Topic and Library/corpus. 5. Marex briefing CU — source-bound synthesis. 6. Call me Will — Directive. 7. Mix harsh at 3kHz — source/project-bounded Assertion or IssueFrame. 8. Pineapple allergy updated — Assertion variants + procedure/directive. 9. Topic extraction over same corpus as Library ingestion — dedupe. 10. RecentActivityRollup — orientation only. 11. Topic Notice vs Topic Slice. 12. Library Notice vs Library Source Slice. 13. Cross-scope/ethical-wall block. 14. Policy drift restamping. 15. False suppression sampling. 16. KDA reference-only rendering. 17. Final-prompt delivery proof before utility. --- ## 20. Open questions for reviewers 1. Is `Assertion` the right canonical term? 2. Should `PremiseFamily` be fully retired or kept as an alias? 3. Does the CU/Assertion boundary work? 4. Does TopicLens + TopicCollectionDirective cover both user value and extraction risk? 5. Is the three-plan coordination model sufficient? 6. Does the context-product registry make injection clear and effective? 7. Does Policy Plane sit at the right level? 8. Is Scope Resolution over-structured or necessary? 9. Are learning targets granular enough? 10. Is DAMS correctly scoped as substrate only?