flowchart LR
E["Evidence-bearing object or relation"] --> A["Lattice admission: identity, ontology, provenance, rights"]
A --> J{"Join, contradict, supersede, or remain plural?"}
J --> S["Versioned semantic state and temporal validity"]
S --> Q["Consumer-scoped retrieval with support and taint"]
Q --> U["Use receipt, contradiction, or correction"]
Q --> R{"Authority, purpose, and freshness still hold?"}
R -- "no" --> X["Refuse, retrieve alternatives, and open residual"]
R -- "yes" --> U
X -. "repair or reauthorization" .-> A
U --> C["Compaction, forgetting, migration, and restart checkpoint"]
C -. "new evidence or ontology change" .-> J
36 Durable Semantic Memory and Knowledge Lattices
36.1 Chapter status
| Field | Value |
|---|---|
| Chapter ID | durable-semantic-memory-and-knowledge-lattices |
| Part | Part II - Planning, Memory, Reasoning, and Execution |
| Status | conceptual |
| Manuscript maturity | v0.3 concept-complete argument-level manuscript |
| Last updated | 2026-08-08 |
| Claim label | Design rationale |
| Evidence level | argument |
| Source loading state | source notes: qcsa_whitepaper, vcm_public, treellm, portia_synapse, ext_graphrag_2024, ext_hipporag_2024, ext_mem0_2025, ext_titans_2025, ext_hermes_agent_2026; raw cache: vcm_public, treellm |
| Test state | The chapter defines a minimum implementation and falsification plan; no chapter-core promotion follows from prose or source synthesis. |
36.2 Drafting guardrail
This chapter owns durable semantic objects and their revision history, not a generic database and not a claim that stored or retrieved content is true. Persistence, graph connectivity, model recall, and fluent reconstruction are storage or access properties until evidence and authority are joined.
36.3 Human Reading Path
Concrete lens. The simpler baseline rebuilds an embedding or graph index and treats successful retrieval as continuity. The chapter requires ontology-loss records and invalidation of every known consumer that used the changed relation.
Context is what a system can see now; durable semantic memory is what it believes it may carry forward. A retrieved statement may be stale, duplicated, contradicted, derived from a denied source, or expressed in an ontology that no longer fits the task. Memory must preserve why content exists, which version is authoritative, what depends on it, and who may revise it.
A governed knowledge lattice keeps assertions, observations, summaries, rules, identities, provenance, confidence, validity, permissions, contradictions, and derivations distinct but linked. Each memory packet records content, evidence ceiling, transformation, scope, retention and deletion duties, and downstream dependencies. Retrieval becomes a bounded proposal subject to relevance, authority, conflict, freshness, and policy checks rather than injection of semantically similar text.
The lifecycle runs from admission and normalization through linking, consolidation, retrieval, use, correction, forgetting, and descendant repair. Summary drift, ontology collapse, poisoned memory, duplicated authority, temporal leakage, incomplete deletion, and synthesis across incompatible contexts remain visible. Durable memory must preserve meaning and obligations across time, while retrieval never proves that the returned content is still true, authorized, or safe to apply.
36.4 Problem
Memory becomes harder as the system improves because more consumers depend on the same objects while their schemas, tasks, and rights diverge. A corrected fact may need to coexist with the historical state that justified an earlier decision; an alias may be valid in one ontology and collide in another; a deletion request may remove use authority without proving that every learned influence or backup copy vanished.
Long-lived AI systems need semantic objects that survive restarts, support graph and associative retrieval, evolve across ontology versions, preserve provenance, merge and retract conflicting assertions, forget under policy, and remain distinguishable from transient context and model weights. Those obligations persist across consumers.
Without a semantic-memory owner, retrieval systems, context stores, claim ledgers, and model parameters can each retain fragments while contradiction, supersession, temporal validity, rights, and restart recovery fall between them. The shared lifecycle method supplies custody; the lattice supplies semantic identity and revision.
36.5 Why existing approaches are insufficient
High recall can hide semantic damage. A vector store may retrieve the right passage under the wrong entity identity; a graph may preserve an edge after its source expires; a long-context model may quote a superseded state; and a summary may compress disagreement into a confident synthesis. The missing property is governed revision across consumers, not another retrieval index.
Vector stores, knowledge graphs, GraphRAG, conversational memory, long-context models, and learned memory can improve retrieval, but none automatically supplies stable object identity, temporal validity, typed relations, provenance-preserving revision, ontology migration, poisoning controls, rights propagation, compaction, and restart-consistent persistence.
Virtual Context, Context Transactions, Claim Ledgers, Artifact Graphs, and ordinary databases are the strongest alternative composition. It wins if their join preserves object and relation identity, ontology migration, support and time, contradiction, rights, compaction, forgetting, restart recovery, and consumer-specific retrieval without a separate lattice.
What this semantic-memory diagram shows: a versioned semantic object with provenance and unresolved alternatives moves. Retrieval does not collapse support, contradiction, rights, or temporal validity.
36.6 Core Claim
Reader claim. Durable memory is not a larger context window. It is a revision system that lets a consumer recover which semantic object it used, under which ontology, evidence, time, rights, and contradiction state.
Operational rule. Admit and retrieve memory only through stable object identity, versioned provenance, temporal and rights checks, explicit contradictions, and a use receipt. A correction must invalidate affected consumers without erasing the historical basis of earlier decisions.
[durable-semantic-memory-and-knowledge-lattices.core, label: Design rationale, support: argument] Durable semantic memory should be admitted through a versioned knowledge-lattice contract that binds object and relation identity, ontology, provenance, support state, temporal validity, authority and rights, merge and supersession, contradiction, retrieval route, compaction and forgetting, restart recovery, consumer use, and residual uncertainty; retrieval quality, graph connectivity, model recall, persistence, or a fluent answer alone establishes neither truth, complete memory, safe consolidation, erasure, nor decision authority.
36.7 Mechanism
36.7.1 Worked revision: a lossy ontology migration reaches its consumers
The finite dossier in tests/fixtures/proof_models/durable_semantic_memory_dossier.json contains two parent objects used by one retrieval. Object 1 migrates exactly. Object 2 does not: its migration is marked lossy, the loss is recorded, and affected consumers are invalidated. Both objects retain provenance and current support, rights, and contradiction checks; the use receipt binds object 1 to ontology version 3, evidence epoch 9, and consumer purpose 5. A retrieval system that merely rebuilds its embedding index would miss the important event: one consumer’s prior semantic basis is no longer safe to reuse under the new ontology.
The local review walks a seven-transition lifecycle and rejects 38 single-axis mutations with exact repair or refusal dispositions. It separately invalidates receipts when object, ontology, evidence epoch, or purpose changes. The test does not show that either stored claim is true or that retrieval is useful. It shows the narrower mechanism the chapter needs: history can remain available while a lossy revision stops silently flowing to a dependent consumer.
The lattice is event-sourced. Admission creates an immutable semantic event plus a current view; later evidence adds contradiction, support, supersession, retraction, or equivalence edges rather than overwriting history. Stable object identity is separated from names and embeddings, so aliases can change and representations can be rebuilt without inventing a new referent. Collision records preserve cases where identity cannot be resolved safely.
Ontology versions are executable migration contracts. A migration names source and destination schemas, exact and approximate mappings, unmapped objects, lossy relations, affected consumers, and a rollback or dual-read window. Consumers bind to an ontology version and receive invalidation when a change alters a relation they used. Plural views are permitted when domains disagree; the lattice need not force one canonical ontology before retrieval can proceed.
Retrieval is planned as a typed query over identity, time, support, rights, and consumer purpose. Exact lookup, vector similarity, graph traversal, temporal filtering, and learned navigation may propose candidates, but the final packet records why each object was selected and which filters denied alternatives. Use receipts then connect a downstream conclusion to the actual memory basis, enabling correction to find affected consumers rather than merely updating the store.
Within the Developmental Intelligence Loop, this chapter owns durable consolidation of admitted semantic objects and their provenance, contradiction, expiry, and forgetting obligations. It receives qualified observations, abstractions, and reality residuals; it does not decide whether a procedure is competent, whether a trained candidate is ready, or whether any memory-derived proposal may act.
Contract. Assign stable semantic identities to entities, events, claims, relations, procedures, and source objects while retaining aliases, uncertainty, and collision records.
Admission. Version ontologies and relation schemas; migrate through explicit mappings that preserve losses, unresolved cases, and invalidated consumers.
Execution. Represent provenance, support, temporal scope, authority, rights, contradictions, supersession, retraction, and derived dependencies on every memory object.
Observation. Combine exact, vector, graph, associative, temporal, and learned navigation under a retrieval plan that records which objects were actually used.
Closure. Consolidate, compact, expire, forget, and recover transactionally, separating storage erasure, retrieval suppression, behavioral forgetting, influence, privacy, and backup state.
36.8 Concept-completion ledger
36.8.1 Semantic identity and provenance
Mechanism. Assign a stable opaque identity to each admitted entity, event, claim, relation, procedure, and source object independently of its name, embedding, graph location, summary, or physical storage address. Every assertion links to the occurrence that produced it, the transformation that normalized it, its ontology version, evidence ceiling, authority, temporal scope, and known aliases or collisions. Derived summaries and learned states remain children, so a representation can be rebuilt without silently changing the referent.
Failure mode. Similarity can merge different people or events, while synonym drift can split one object into incompatible duplicates. A popular summary can become the apparent source after its parents expire, and an address migration can retarget consumers to a different object while preserving a familiar label.
Non-claim. Stable identity and complete provenance do not establish that the object exists as described, that an assertion is true, or that every real-world referent has been uniquely resolved.
Source grounding. qcsa_whitepaper supplies the separation among semantic object identity, plural virtual addresses, evidence-bearing graph state, and physical routes. ext_graphrag_2024 and ext_hipporag_2024 provide graph-derived retrieval comparators whose generated structures remain fallible. No source establishes local identity accuracy or truth.
36.8.3 Consolidation and compaction
Mechanism. Consolidate related events only through a versioned transformation that lists covered parents, retained disagreements, omitted detail, temporal assumptions, rights intersection, reconstruction route, and consumers allowed to use the derivative. Compaction may replace hot representations or indexes for efficiency, but the durable event lineage and invalidation edges remain. Learned memory checkpoints carry training inputs, update rule, optimizer state where relevant, and a route back to exact records rather than becoming an uninspectable canonical truth.
Failure mode. Summaries can smooth contradictions, community labels can invent consensus, and compression can discard the detail needed to correct a downstream decision. A snapshot may preserve semantic content while losing deletion duties or may resurrect retracted relations after recovery. Keeping every byte forever creates a different failure through cost and privacy exposure.
Non-claim. A compact representation that answers a benchmark correctly is not necessarily faithful, complete, reversible, private, or safe for every consumer.
Source grounding. ext_graphrag_2024 motivates graph and community-summary derivatives; ext_mem0_2025 motivates consolidation under latency and token constraints; ext_titans_2025 motivates mutable learned memory. qcsa_whitepaper supplies residual- and migration-aware compression boundaries. None proves lossless consolidation here.
36.8.4 Contradiction and belief revision
Mechanism. Represent contradiction, support, supersession, retraction, equivalence, and unresolved plurality as typed edges rather than overwriting a current value. Temporal validity separates “was true then” from “is supported now,” and a claim-ledger interface determines belief state while memory preserves the underlying objects. Consumers request a purpose- and time-bounded view that includes material alternatives and receive invalidation when later evidence changes an object they used.
Failure mode. Last-write-wins can erase history; majority retrieval can suppress a rare correct correction; and a generated synthesis can combine mutually exclusive claims into a fluent false compromise. The opposite failure sends every historical contradiction to every consumer, overwhelming useful retrieval and encouraging them to ignore provenance.
Non-claim. Preserving disagreement does not decide which assertion is true, guarantee calibrated belief revision, or make a contradiction resolvable with available evidence.
Source grounding. qcsa_whitepaper explicitly separates propositions, evidence, provenance, belief state, contradiction, and permitted use. ext_graphrag_2024 and ext_hipporag_2024 motivate graph traversal but do not supply governed belief revision. The chapter’s event-sourced revision design remains an argument, not a reproduced result.
36.8.5 Identity levels and plural atlas geometry
QCSA prevents durable memory from collapsing four different objects into one vector key. An occurrence is the exact mention, image region, sensor span, or code reference in context. A type is a reusable abstraction. An instance is one world or system object. An expression composes objects, roles, negation, quantity, time, modality, causality, and operators. The word “bank,” a river-edge type, one particular bank, and “the fisherman sat on the bank yesterday” require different identities and lifecycle rules. Similarity, translation, synonymy, and shared function remain evidence for identity—not identity itself.
The address atlas is consequently plural. Each facet declares its consumer, optimization target, relation and grounding sources, branching/depth policy, calibration, migration cadence, and prohibited uses. Paths may be soft, variable-depth, mixed-arity, and multi-parent. A boundary object can retain several candidate paths until a consumer-specific threshold requires commitment. Branching width is chosen against joint compute, serial depth, classification error, load balance, migration stability, accelerator fit, and semantic-purity cost rather than a universal binary rule.
Plural facets also provide error-detecting redundancy. An ontological path suggesting a financial institution while contextual and linguistic facets suggest a river edge triggers re-resolution rather than a tidy but wrong record. Cross-facet disagreement is not proof that one facet is false; it is a typed reason to preserve alternatives, acquire evidence, or abstain. The atlas epoch publishes topology, intended consumers, objectives, representative and boundary cases, calibration, utilization, migration maps, known collisions, blind spots, prohibited uses, signatures, and rollback references. This makes semantic organization a governed artifact rather than an invisible side effect of embedding training.
36.8.6 TreeLLM after its own correction history
The complete TreeLLM lineage makes the knowledge-lattice idea more useful by showing how it fails when stated too strongly. Its first paper separates a traversable semantic graph from a learned navigator and gives model-visible symbols paths, types, probabilities, and residual attributes. The next drafts turn thirteen English questions into eternal roots, declare one global graph the source of truth, freeze the navigator forever, and attach perfect grounding and device-scale claims. A later internal critique rejects those choices; revised versions replace fixed roots with learned or evolvable anchors and add approximate fallbacks, private overlays, streaming ingestion, semantic compaction, bounded graph snapshots, caching, and speculative prefetch. The correction sequence is more informative than any “final” label: every convenient hidden assumption eventually returns as an identity, version, authority, bandwidth, ambiguity, update, or governance obligation.
The stack retains that lesson through four separations. First, a stable object or relation identity is not its HLSH coordinate, path, embedding, canonical text, anchor, or storage key. Those are versioned atlas representations that may migrate. Second, a graph record is not a belief verdict. Observation, quotation, user assertion, model inference, rule, contradiction, and retraction enter as different objects with different authority. Third, a retrieval route is not evidence acceptance. Exact lookup, graph traversal, vector fallback, learned navigation, exploratory bridging, and external tool acquisition are separately typed proposal routes. Fourth, a representation’s fixed byte width is not its semantic adequacy. A compact token needs a source or richer-representation fallback, collision and ambiguity state, graph and codec epoch, task envelope, expiry, and a consumer-specific residual.
TreeLLM’s Synapse, Navigator, Scout, and Ingestor names can therefore be reconstructed as an ordinary routing contract. The exact route retrieves a known identity or typed edge. The approximate route uses vector or ANN search when hard structure is insufficient. The exploratory route proposes a bridge without writing it as fact. The ingestion route proposes objects, aliases, relations, and evidence packets. A tool route obtains new observations through the Runtime Adapter owner. An abstain or escalation route handles unresolved identity, support, rights, or budget. Route confidence is calibrated against route-specific error; it cannot authorize persistence or harden an edge. Repeated traversal is popularity, not corroboration.
The later DklSnapshot sketch also supplies a useful bounded alternative to “infinite context.” A snapshot request names entry objects, graph epoch, hop depth, node budget, allowed relation types, weight threshold, time, rights, purpose, and consumer. The result includes admitted nodes, the routes that found them, denied candidates, provenance, contradictions, omissions, expiry, and the objects actually delivered and used. Writing an old conversation as a session-summary node is a lossy derivative with parents and residuals, not exact recall. Likewise, an HLSH prefix or graph path can help retrieve a candidate without proving that the consumer saw, used, or correctly interpreted it.
An update is effect-complete only when its descendants close. A single node write may be constant-time at the database API while entity resolution, sense separation, contradiction, rights, relation validation, address/token migration, index and cache invalidation, snapshot expiry, learned-router calibration, backup state, deletion duties, and consumer notices are not. Append-only event history can preserve how a view changed, but it does not excuse retraction, redaction, or erasure. “O(1) update” is therefore reserved for the narrow storage operation unless the complete semantic closure is measured.
TreeLLM calls background graph repair the Gardener. Here it is a proposal engine, not a semantic sovereign. A merge of close nodes records stable parent identities, sense-aware preconditions, affected relations and consumers, retained disagreement, rights intersection, before/after adequacy, reversible migration, and rollback. Soft edges found by approximate search remain soft until independent evidence and admission policy justify a typed relation. Otherwise a popular false path can become permanent through its own retrieval feedback.
Private user overlays, organizational stores, public evidence, and federated contributions remain different tiers. A read does not simply union them: precedence, conflict, authority, temporal scope, and disclosure depend on the consumer. Each tier carries fork, correction, export, deletion, and audit rights. Federated graph or gradient updates need a threat model, consent, poisoning and Sybil controls, rollback, and privacy accounting; keeping a mobile copy or a shared Rust core does not by itself make the system sovereign.
Finally, adaptive token caching and speculative traversal remain performance hypotheses behind the same contract. A short id is scoped to a session, model, token codec, and graph epoch; it is collision checked, invalidated on update, bounded against thrashing, and falls back to the full representation. Speculative paths may prefetch several candidates, but wrong-path reads, privacy crossings, stale epochs, wasted bandwidth, and tail latency remain in the cost record. Only committed and verified hops may appear in a downstream route receipt. The source-reported percentages and production claims supply no local evidence.
Source grounding. treellm supplies the correction lineage, structured token variants, DKL records, router roles, exact/soft traversal, Ingestor, Gardener, overlays, snapshots, skill subgraphs, caching, speculation, and mobile/federated proposals. Its 65,779-word cache contains no inspected local implementation or reproduced result, and four final tabs are Aletheia source contamination rather than independent support. The mechanism above is a conservative reconstruction at argument support.
PortiaSynapse makes the learned-navigator boundary more concrete without changing the authority model. Its RichContext joins coordinate bits, content, edge distribution, neighbor summary, and modulation into one input packet. That packet must bind the exact DKL snapshot, codec, neighbor-selection policy, provenance, rights filter, and construction version; otherwise the same vector shape can describe different worlds. Its mutable Focus memory is request- and session-scoped, resettable, replayable, and excluded from cross-user reuse by default. A lock can make mutation race-safe, but it cannot make state semantically isolated or deletion-complete.
36.8.7 Retrieval and context assembly
Mechanism. Compile retrieval as a plan over exact identity, semantic similarity, graph relations, time, rights, support, task purpose, and cost. Each candidate carries the route that found it and each denial records the failed filter. Context assembly selects a bounded set with provenance, contradictions, expiry, and residual uncertainty, then records which objects the model actually received and which were cited or used downstream. Learned navigation remains a replaceable proposal mechanism behind this contract.
Failure mode. Vector similarity can return a plausible but wrong entity; graph centrality can amplify popular misinformation; a long context can contain the answer without the model using it; and a global summary can omit local evidence. Evaluation can hide these failures by using static corpora, LLM judges, or queries that saturate across all methods.
Non-claim. Retrieval recall, answer accuracy, graph connectivity, or context length does not prove durable memory truth, causal use, safe authority, or general task adequacy.
Source grounding. ext_graphrag_2024, ext_hipporag_2024, and ext_mem0_2025 provide distinct graph, associative, and conversational retrieval comparators. qcsa_whitepaper contributes active, consumer-scoped semantic addressing. Their reported results are not local reproductions and do not establish this joined plan’s advantage.
36.8.9 Retention, forgetting, and deletion
Mechanism. Model retention as separate authorities over storage, retrieval, use, training influence, behavioral expression, privacy, backups, and legal or audit preservation. A deletion event identifies the object, purpose, affected derivatives, indexes, caches, checkpoints, backups, and consumers; each returns a disposition and residual. Tombstones and minimal accountability records may remain where authorized so a forbidden object is not silently reintroduced, while content access is removed according to the governing duty.
A learned retention or compression policy also freezes a decision-time observable manifest. It may use only the events, features, page boundaries, source roles, policy state, and costs actually available when the choice was made. Later queries, answers, gold evidence labels, benchmark dependency annotations, and post-decision page boundaries attach to a separate hindsight record for learning and counterfactual analysis; they do not rewrite the original manifest. A future-aware oracle is reported as an upper bound, never as the deployable policy.
This barrier matters because a memory system can appear prescient when its summaries, retention labels, page boundaries, or value model were built with knowledge of the later question. Evaluation must replay the decision from the frozen observable record, report blind utility and oracle gap separately, and include delayed utility, distribution shift, costly reacquisition, stale evidence, and resurfacing of rarely used but consequential pages. Low past retrieval frequency alone cannot authorize permanent deletion.
Failure mode. Deleting one database row can leave embeddings, summaries, caches, model influence, and backups active. Conversely, claiming complete erasure may conflict with audit obligations, while an overbroad tombstone can reveal that sensitive content existed. Behavioral absence on a probe can be mistaken for influence or storage erasure.
Non-claim. Retrieval suppression does not prove storage deletion, model unlearning, privacy repair, causal influence removal, or irrecoverability from every descendant.
Source grounding. ext_mem0_2025 makes memory extraction and consolidation operationally salient but does not establish deletion completeness. ext_titans_2025 raises mutable neural-memory custody and rollback duties. qcsa_whitepaper supplies lifecycle and descendant concepts while explicitly not proving forgetting or erasure.
36.8.10 Poisoning, taint, and semantic quarantine
Mechanism. Attach taint to a source, transformation, object, relation, index, or learned checkpoint and propagate it through known dependencies. Admission tests inject alias collisions, adversarial relations, stale epochs, forged provenance, and repeated self-citation. Suspect objects move to a quarantine view where authorized reviewers can compare clean parents, re-resolve identities, rebuild indexes, and issue descendant invalidations without deleting the forensic trail.
Failure mode. A poisoned memory can become more trusted through repetition, graph centrality, or model-generated summaries. Rebuilding an index from tainted snapshots can resurrect the attack; indiscriminate taint propagation can quarantine most of the store and create denial of service. A detector tuned to obvious corruptions can give false confidence against subtle semantic poisoning.
Non-claim. Passing a finite poisoning suite does not prove adversarial robustness, source honesty, or safe operation against novel attacks.
Source grounding. qcsa_whitepaper identifies alias-driven privilege escalation, atlas poisoning, route disagreement, and stale-epoch failures. ext_graphrag_2024, ext_hipporag_2024, and ext_mem0_2025 motivate derived graph and retrieval layers whose poisoning behavior is not established by their source notes. No local defense efficacy is claimed.
36.8.11 Lineage, rollback, restart, and descendant repair
Mechanism. Checkpoint the event log, current semantic view, indexes, ontology mappings, rights state, learned-memory state, caches, RNG or update state where applicable, and outstanding consumer receipts under one authority. Restart replay must either reproduce the governed view or enumerate every nondeterministic and lossy difference. A correction walks the descendant graph to summaries, embeddings, contexts, plans, and decisions; rollback restores an authorized consistent state and replays later admissible events.
Failure mode. Restoring only the database can leave stale indexes and caches active, while restoring a snapshot can resurrect deleted content or old rights. A syntactically successful rollback may not repair downstream decisions already made. Replaying a learned memory without optimizer or update state can produce an apparently equivalent but semantically different system.
Non-claim. Byte-equivalent restart does not establish semantic truth, and administrative rollback does not reverse external effects or prove every descendant repaired.
Source grounding. qcsa_whitepaper supplies immutable epochs, migration records, compatibility, rollback, and explicit residuals. ext_titans_2025 motivates custody of mutable neural state. The other retrieval comparators do not establish restart equivalence; the proposed full-state repair path remains untested.
36.9 Interfaces
Writes and reads cross different authorities. Evidence and Claim Ledgers may propose semantic events; privacy and data-rights owners constrain retention and use; Context Transactions obtain isolated views; consumers return use and error receipts. The lattice can preserve and retrieve these objects, but it cannot promote a proposition, waive a right, or decide that a procedure is competent.
Virtual Context assembles a temporary working set; Context Transactions own snapshots and taint; Claim Ledgers own proposition belief state; Artifact Graphs own execution lineage. The lattice owns durable object/relation identity, ontology evolution, semantic joins, and retrieval across restarts.
- Virtual Context ABI materializes bounded consumer packets; this chapter owns the durable semantic substrate it reads.
- Context Transactions owns isolation, commit, mounts, taint, and crash semantics across state changes.
- Claim Ledgers owns belief support and revision; durable memory stores the semantic objects and relations those claims reference.
- Procedural Memory owns reusable action trajectories; Artifact Graphs owns generic evidence lineage.
- Privacy/Data Rights and Data Engines govern rights, deletion, learned influence, and descendant obligations.
36.10 Invariants
Derived summaries, embeddings, community labels, and learned memory states are children of source objects, not replacements for them. Their generation method, coverage, ontology, and invalidation edge remain queryable. A consumer may use a derivative for efficiency only while the parent evidence and rights contract permits that use.
Applying the shared lifecycle method, every stored relation retains provenance, support, time, rights, and revision lineage; contradiction stays representable; compaction and forgetting cannot silently rewrite a consumer’s past basis.
- Object identity, source identity, semantic equivalence, and aliasing remain distinct.
- No merge erases provenance, contradiction, uncertainty, temporal scope, or rights.
- Ontology migration records unmapped and lossy cases and invalidates affected consumers.
- Retrieval records actual use; storage presence does not imply influence or belief.
- Restart recovery, compaction, forgetting, deletion, and model unlearning remain separate claims.
36.11 Formalization hooks
AsiStackProofs.DurableSemanticMemoryReview contributes 35 theorem declarations over a reachable seven-transition review. A complete authored dossier reaches only eligibility for a Project Theseus memory replay and retrieval campaign. Each of 38 admission-axis mutations rejects readiness and receives an exact repair or refusal disposition.
The formal model keeps representation, alias, and semantic-object identity separate. Induction over arbitrary finite parent sets proves that collected provenance retains every parent source and that a derived use authorized by all parents cannot exceed any one parent’s purpose scope. Lossy ontology migration without affected-consumer invalidation is rejected. Retrieval use receipts bind every used object to a current, rights-checked, contradiction-aware candidate, and event-log replay composes exactly across list concatenation.
Object, ontology, evidence-epoch, and consumer-purpose changes invalidate a memory receipt. Two collision pairs show that identical summary signals can hide opposite contradiction states and identical storage-deletion signals can hide opposite learned-influence states. A consumer bridge sends an open memory deletion duty into Context Transactions, which blocks materialization without closure or declassification. All dossier fields are authored assumptions. No theorem establishes semantic truth, useful retrieval, complete memory, behavioral forgetting, semantic completeness after restart, deployment, support, transfer, or external effect. Chapter support remains argument. The model does not prove that any stored object is true or that any retrieval, migration, deletion, replay, or repair mechanism works in a deployed system.
36.12 Failure modes
Two operational failures are especially deceptive. Semantic resurrection occurs when a compacted snapshot, backup, cache, or embedding index restores a retracted relation after restart. Authority laundering occurs when a highly ranked memory object enters a prompt as if retrieval rank implied truth or permission. Recovery and consumer-bound admission tests must detect both.
The principal failure family includes entity collision; duplicate identity; stale truth; ontology drift; relation poisoning; provenance loss; contradiction collapse; false supersession; retrieval popularity bias; privacy leakage; compaction damage; incomplete forgetting; crash inconsistency; backup resurrection; memory-to-context authority laundering.
Evaluation must distinguish retrieval failure from absent storage, identity collision, stale ontology, unsupported merge, rights denial, and reconstruction error. Positive controls seed contradictions, superseded facts, expired relations, and restart faults so a smooth answer cannot hide semantic corruption.
36.13 Minimum Viable Implementation
The implementation needs an observable replay path. Starting from an empty store, the same admitted event stream should rebuild the same governed current view, or report every nondeterministic and lossy difference. Crash injection occurs before and after event append, index update, snapshot creation, compaction, rights change, and consumer receipt. This establishes transactional behavior; semantic truth remains outside the receipt.
Implement an event-sourced semantic-object store with typed nodes and relations, temporal validity, provenance, contradiction, supersession, ontology version, rights, and transactional snapshots. Compare exact, vector, graph, and hybrid retrieval on update-heavy tasks with injected collisions, stale facts, conflicting sources, poisoning, deletion, compaction, crash, and restart. Measure task utility together with provenance survival, contradiction calibration, rights closure, latency, and residuals.
The minimum lattice stores a bounded typed corpus with conflicting and time-varying relations, explicit provenance and rights, consumer-specific retrieval, restart checkpoints, compaction, and deletion requests. It compares an ordinary vector/graph store with the governed lattice on correction and recovery as well as retrieval.
36.14 Evidence and falsification program
Argument exit requires held-out retrieval and update campaigns spanning contradiction, temporal change, ontology migration, restart, compaction, rights, and deletion. Strong database, vector-search, graph, and claim-ledger compositions are matched on information access and tuning; correctness, calibration, latency, storage cost, and residuals are scored independently.
36.15 Mature Research Target
The mature lattice is a shared semantic substrate for heterogeneous models, tools, agents, and institutions. It supports several ontologies and uncertainty representations at once, negotiates mappings at the interface, and allows a consumer to request a bounded view without copying the entire knowledge state. Hot indexes and learned navigators remain replaceable accelerators over a durable event and identity layer.
Evaluation uses update-heavy natural workloads rather than static question answering alone. It measures correction propagation, temporal reasoning, contradiction preservation, ontology migration, rights enforcement, restart equivalence, poisoning recovery, and downstream task utility against strong database, vector, graph, and long-context compositions. Storage, latency, energy, reviewer burden, and unresolved collision rates remain joint costs.
The strongest outcome is not perfect memory. It is controlled fallibility: the system can state what object it used, how that object changed, which consumers may rely on it, what disagreement survives, and what could not be forgotten or reconstructed. That history lets the wider stack correct decisions without pretending one embedding space or summary is the world.
The mature lattice lets many cognitive substrates share durable meaning without pretending that one embedding space or ontology is final. It supports plural hypotheses, explicit conflict, reversible compaction, rights-aware forgetting, restart recovery, and consumer-bound retrieval while preserving the history needed to correct itself.
Current evidence does not establish that endpoint. Memory-lattice support remains at argument until longitudinal replay demonstrates identity continuity, contradiction handling, provenance-complete revision, descendant invalidation, and bounded retrieval across model replacement.
36.16 Codex test plan
| Test | Purpose | Status |
|---|---|---|
| Identity collision | Present near-duplicate objects with conflicting provenance and reject an unsupported merge. | planned |
| Temporal contradiction | Preserve old and new states, validity intervals, and the consumer basis instead of overwriting history. | planned |
| Restart and compaction | Replay from checkpoint after compaction and expose every lost or reconstructed relation as a residual. | planned |
| Rights-aware retrieval | Deny an unauthorized consumer without deleting the semantic object’s existence or revision record. | planned |
Planned formalization route: lean:durable-semantic-memory-and-knowledge-lattices.admission_boundary covers only the finite admission record for durable semantic memory and knowledge lattices; it cannot establish whether durable semantic memory and knowledge lattices works safely or transfers.
36.17 Memory as a Persistence Locus
A memory write is a durable behavioral intervention, not clerical storage. The Adaptive Commit Boundary treats memory as one candidate locus among several and requires the write to retain its lesson identity, evidence, scope, commitment profile, authority, qualification conditions, expiry, monitor, and descendant links. A fact that belongs in a dated memory may not belong in weights; a procedural defect may require a test or tool rather than another retrieval.
Deletion must follow descendants and reliance. Removing one source node does not erase summaries, compiled context, cached routes, tools, or institutional decisions derived from it. Anything that cannot be removed, deoptimized, compensated, or requalified remains an explicit residual.
36.18 Source crosswalk
| Source ID | Title | Bounded use |
|---|---|---|
vcm_public |
Virtual Context Memory | Passage-reviewed Corben architecture for immutable events and semantic pages, representation graphs, contradiction/rejection memory, decision-time observable retention features, anti-starvation, invalidation, deletion closure, and blind-policy versus future-aware-oracle separation. Its activated-fault study and uninformative packet pilot do not establish durable-memory truth, retention-policy quality, natural outcomes, deletion/unlearning completeness, privacy, or chapter-core support. |
qcsa_whitepaper |
Question-Compiled Semantic Addressing | Corben-authored semantic-control-plane lineage for stable object identity, versioned addresses, evidence-bearing hypergraphs, lifecycle migration, and explicit residuals. Its bounded local artifacts do not establish durable memory truth, deletion completeness, restart equivalence, retrieval quality, or production transfer. |
treellm |
TreeLLM correction lineage | Complete 28-tab visible lineage for explicit semantic graphs, structured tokens, model/memory separation, exact/approximate/exploratory/ingestion routing, learned anchors, overlays, snapshots, graph hygiene, adaptive caching, speculative traversal, skills, and mobile/federated variants. The source’s own critique controls frozen-root and final-architecture drafts; no local implementation, graph, token, HLSH property, benchmark, O(1) semantic update, grounding, privacy, or production result is established. |
portia_synapse |
PortiaSynapse | Authenticated successor paper that instantiates a DKL navigator over a 512-dimensional RichContext and mutable Focus memory. It motivates snapshot/codec/provenance binding and request-scoped state; its source-reported implementation does not establish memory truth, isolation, deletion closure, learning, or routing quality. |
ext_graphrag_2024 |
From Local to Global: A Graph RAG Approach to Query-Focused Summarization | Primary GraphRAG comparator for LLM-derived entity graphs, community summaries, and global corpus questions; generated graph and summary layers remain fallible derived representations and do not establish truth, complete coverage, local adequacy, or an ASI Stack memory result. |
ext_hipporag_2024 |
HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models | Primary NeurIPS comparator for knowledge-graph retrieval with Personalized PageRank and single-step associative navigation; reported multi-hop QA gains do not establish durable truth, update correctness, resistance to poisoning, local reproduction, or a general memory system. |
ext_mem0_2025 |
Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory | Primary preprint comparator for extracting, consolidating, retrieving, and graph-linking conversational memory under latency and token-cost constraints; LOCOMO and LLM-judge results do not establish fact correctness, poisoning resistance, general memory, local reproduction, or production readiness here. |
ext_titans_2025 |
Titans: Learning to Memorize at Test Time | Primary neural-memory comparator for test-time memorization and long-context sequence modeling. The paper motivates mutable-state provenance and rollback tests; no local model or benchmark result is reproduced. |
36.18.1 Manifest source assignment reconciliation
These rows keep Durable Semantic Memory and Knowledge Lattices’s manifest assignments visible at their recorded review boundary. Passage review does not establish local reproduction, performance, safety, deployment, or support-state movement.
| Source | Intake role | Boundary |
|---|---|---|
ext_hermes_agent_2026 |
Passage-reviewed comparator: Hermes Agent: Learning, Memory, Tools, and Security Architecture. Supplies a current implementation comparator that separates bounded always-present MEMORY/USER snapshots, on-demand FTS5 session search, and longer load-on-demand procedural skills, while scanning admitted prompt memory for injection and exfiltration patterns. | Official software and documentation review only; no memory fact, search result, poisoning control, session isolation, learning behavior, cost, utility, or deployment result was reproduced. No local implementation, reproduction, performance, safety, deployment, support-state, or ASI result is established by this reconciliation row. |
adjudicated_persistence |
Passage-reviewed comparator: Adjudicated Persistence: Governing the Transition from Experience to Durable Structure in Adaptive Systems. Treats memory as one consequential persistence locus whose writes require evidence, authority, qualification, expiry, and descendant-aware revocation. | Conceptual author framework and benchmark proposal; no local implementation, empirical result, independently checked proof, safety result, or support movement. No local implementation, reproduction, performance, safety, deployment, support-state, or ASI result is established by this reconciliation row. |
36.19 Summary
Durable semantic memory separates referents from names, evidence from derived representations, storage from belief, retrieval from authority, and suppression from erasure. Event-sourced identity and revision records preserve conflicts, temporal validity, provenance, rights, and ontology changes across restarts and many consumers.
The knowledge lattice is therefore not a claim of total recall. It is a governed service for admitting, relating, finding, using, correcting, compacting, and forgetting semantic objects while exposing losses and residuals. Its value is measured in correction and recovery as well as retrieval quality.
Durable semantic memory is a revision-capable knowledge service, not a bigger prompt. It preserves object and relation identity, ontology, provenance, support, time, rights, contradiction, retrieval, compaction, forgetting, and restart history so downstream cognition can know both what it received and why it may be wrong.
36.20 Handoff
Context Transactions, Snapshots, Mounts, and Taint mounts a consumer-scoped lattice view with exact identities, support, temporal validity, rights, contradictions, and retrieval residuals. A mount does not promote those objects, erase alternatives, authorize writes back into durable memory, or hide which ontology and snapshot produced the view.