Memory and knowledge is dated, sourced memory for compliance leads, so that any answer can be read as of the day it mattered, with later changes shown apart.
Preserve source-connected history with two distinct time axes.
Original or changed policy with source and both time axes.
Preserve source-connected history with two distinct time axes.
Source, holder, validity and recorded-at time.
Earlier records and operation ancestry.
Valid time and recorded-at time remain distinct.
Apply the actual semantics of revision, contest and historical use.
Claims and typed operations.
Validity time, knowledge time and scope.
Apply the actual semantics of revision, contest and historical use.
Supersession, contest, suppression and retraction differ.
Validity question and knowledge cutoff.
Supersession, contest, suppression and retraction have different historical meanings.
Assemble relevant evidence without confusing derived lookup with canonical memory.
Historically appropriate or contested evidence.
Rebuilt derived structures; canonical history remains.
Assemble relevant evidence without confusing derived lookup with canonical memory.
Derived lookup and canonical-source link.
Scope/byte bounds and explicit abstention.
A derived lookup does not replace canonical source memory.
Relate the selected evidence to the right responsibility and objective.
Bounded source-bearing selection or abstention.
Selected role/objective and explicit corrections.
Relate the selected evidence to the right responsibility and objective.
Selected identity, role, infrastructure and purpose.
Corrected inference with its lineage.
A corrected inference retains its source and responsibility context.
Annotate, test and retain useful derived structures without rewriting history.
Canonical records and typed retention requirements.
Measured harms, utility and remaining uncertainty.
Annotate, test and retain useful derived structures without rewriting history.
Duplicates, conflict and typed retention/forgetting.
Proposed generalization remains distinct from promoted instruction.
A proposed generalization does not silently become an instruction.
Keep actual utility, failures and subsequent-use evidence available.
Actual retrieval use, not automatic future utility.
Independent useful/failing reuse evidence.
Keep actual utility, failures and subsequent-use evidence available.
Measured failure and caution/release basis.
Attributed retrieval use and independent future utility.
Claims of future usefulness need independent later-use evidence.
Original or changed policy with source and both time axes.
Preserve source-connected history with two distinct time axes.
Source, holder, validity and recorded-at time.
Earlier records and operation ancestry.
Valid time and recorded-at time remain distinct.
Claims and typed operations.
Validity time, knowledge time and scope.
Apply the actual semantics of revision, contest and historical use.
Supersession, contest, suppression and retraction differ.
Validity question and knowledge cutoff.
Supersession, contest, suppression and retraction have different historical meanings.
Historically appropriate or contested evidence.
Rebuilt derived structures; canonical history remains.
Assemble relevant evidence without confusing derived lookup with canonical memory.
Derived lookup and canonical-source link.
Scope/byte bounds and explicit abstention.
A derived lookup does not replace canonical source memory.
Bounded source-bearing selection or abstention.
Selected role/objective and explicit corrections.
Relate the selected evidence to the right responsibility and objective.
Selected identity, role, infrastructure and purpose.
Corrected inference with its lineage.
A corrected inference retains its source and responsibility context.
Canonical records and typed retention requirements.
Measured harms, utility and remaining uncertainty.
Annotate, test and retain useful derived structures without rewriting history.
Duplicates, conflict and typed retention/forgetting.
Proposed generalization remains distinct from promoted instruction.
A proposed generalization does not silently become an instruction.
Actual retrieval use, not automatic future utility.
Independent useful/failing reuse evidence.
Keep actual utility, failures and subsequent-use evidence available.
Measured failure and caution/release basis.
Attributed retrieval use and independent future utility.
Claims of future usefulness need independent later-use evidence.
An illustrative method for retaining canonical evidence and answering scoped historical questions.
Additional mechanisms are described in the technical reference.
What the research says
The other half is whether it applied on the day in question.
Teams report spending a quarter of the workweek searching, and commercial legal tools built on retrieval still gave wrong answers 17% to 33% of the time. Retrieval finds a document; it says nothing about whether that document applied on the day in question. The memory layer stores when each claim applied and when it was recorded, so a compliance lead can read the answer as of a date and see later changes apart.12
An institution changes a policy on Thursday. A reviewer needs the rule that applied on Monday and what was known when Monday’s decision was made. The next task needs the current rule. A single current document or a plausible summary cannot answer all three questions.
Our knowledge architecture separates source records, claims, their validity and the time they were observed. A dedicated world model connects those records to identities, responsibilities, objectives and the systems a person or organization chooses to connect.
Vagus resolves time and authority. Agency Engine maintains product and organization context through versioned memory operations. The Vault carries methods, assumptions and evidence into reuse. This gives an intelligence environment more structure than an archive of conversation text.
Vagus’s claim algebra separates what a statement says from who supplied it, when it applies and what operations later changed its status.
Validity time answers when a claim applied. Observation time answers when it entered the record. An as-of query can preserve the earlier state of knowledge.
Claims and operations with timestamps.
The applicable evidence at the requested time basis.
The effective authority of a derived claim is bounded by the minimum authority along its contributing ancestry. A missing ancestor or cycle cannot create stronger evidence.
Claims of equal scope and authority that disagree can remain contested. Supersede, revise, contest, suppress and retract carry different meanings.
Suppressing a claim from ordinary recall is distinct from changing its validity or retracting it. The canonical operation record retains the distinction.
The compliance lead asks with a date; knowledge is read as of that date, and later changes are shown apart. Illustrative.
Illustrative situations, connected to the mechanisms that make them possible.
Institutions
A policy changes, but earlier claims and the reasons for past decisions still matter.
R&D and software teams
A new problem may overlap earlier engineering in another project.
Founders and small teams
A prospect moves between a public page, a question and a private conversation.
The memory store uses holder-scoped, append-only records with keyed integrity chains. Its retrieval index is a derived view rather than a second editable history.
A SQLite FTS5 index combines claim, lesson, failure and episode pointers. Rebuild writes a temporary index and swaps it into place atomically.
Verified canonical records.
A replaceable retrieval index.
Recall limits both items and bytes. Ranking combines textual relevance, authority, scope and use-based activation; a result retains its matched evidence and source pointer.
A query, scope and output budget.
Evidence-bearing items, withheld items or an abstention reason.
Readers use the verified prefix of the record chain and retain a finding for unverified material. A broken tail cannot silently become trusted context.
Holder boundaries travel with storage and retrieval. A useful match does not grant permission to disclose the material or act on it.
Agency Engine separates accumulation from interpretation. Its per-origin set merge preserves the information needed to distinguish a changed fact, an incorrect belief and a disagreement between sources.
The data layer merges an observed-remove set by union. The element identity contains both the claim and its origin, so two sources asserting the same fact remain two attributable elements.
Claim/origin elements and their memory operations.
A convergent accumulated set with source identity retained.
A pure read-time function interprets the canonically ordered set under scope, authority, temporal validity and conflict rules. A revised interpretation can rebuild the projection without rewriting the underlying history.
Decayed, never-captured, unlearned and superseded describe different events. Memory maintenance can nominate a change while preserving the historical source and its approved validity operations.
A recurring cluster can justify proposing a generalization. Promotion, retrieval evidence and retention have separate contracts; repeated access inside one concentrated session does not become independent spaced evidence.
The inspected thresholds are declared policy choices. Future utility and retention are evaluated for the domain.
Agency Engine separates recall, proposed generalization, memory maintenance and typed forgetting. Vagus’s failure memory and the Vault’s evidence contracts retain different kinds of consequence.
An episode or attributed claim retains the source behind a decision, correction or failure. Repeated copies do not become independent supporting origins.
Duplicate detection, conflict annotations, decay nomination and approved validity changes are explicit operations. A tidy summary cannot silently replace the source history.
Vagus records defined harms with decay and hysteresis. Restrictive failure memory can keep a known failing context from being treated as a fresh unexamined choice.
A Vault method keeps its contracts and evidence lineage. Unresolved composition assumptions remain visible to the next application.
Proprietary operating knowledge can improve the context for future work. Updating a model is a separate intervention with a different data and evaluation contract.
A corrected policy or preference changes the records and context used for future tasks. The underlying foundation-model weights can remain fixed.
Recorded outcomes can change a task estimate, selection policy or approved workflow. Their acceptance basis is specific to the recurring task.
A trainable model can be studied using permitted examples and held-out future work. Retention, exposure, cost and recovery belong to that study.
Inputs, retained state, outputs and limits of Naevum’s Memory & knowledge systems architecture.
No. It is a scoped representation of connected sources, explicit preferences and supported inferences. Its contents and limits must remain inspectable and correctable.
Connected sources become structured context. Observations keep their origin; inferences remain open to correction. This is the architecture of a dedicated engagement, shaped around the world you choose to connect.
Keep people, roles and delegation sources distinguishable.
Selected roles, accounts and responsibility sources.
Keep people, roles and delegation sources distinguishable.
Selected people, accounts and roles.
Delegation and responsibility sources, not inferred authority.
A relationship or inferred role does not create authority.
Make inclusion, exclusion, access and freshness explicit.
Actual source inclusion and access; not a UI inspection event.
Independent dated observation.
Make inclusion, exclusion, access and freshness explicit.
Explicitly included sources and exclusions.
Dated observation, source and access state.
Opening an explanatory panel does not include a source or grant access.
Retain facts, claims, decisions and corrected interpretations with their lineage.
Source-bearing observation with time and scope.
Selected relevant role and responsibility.
Separately supplied evidence correcting an earlier inference.
Attributed action/consequence history and retained uncertainty.
Retain facts, claims, decisions and corrected interpretations with their lineage.
Evidence, time and unresolved contradiction.
Current inference differs from source observation; correction preserves ancestry.
An inference remains distinguishable from its source observation.
Relate selected assets and dependencies to observations and responsible operators.
Selected assets, capabilities and dependencies.
Relate selected assets and dependencies to observations and responsible operators.
Asset, dependency and operator relationships.
Known capability, unavailable integration and unknown state.
A known capability or dependency is not an execution grant.
Relate context to declared aims while keeping forecasts and inferred habits distinct.
Known capabilities and operator/dependency relationships; no permission.
Explicit objective, commitment or corrected preference.
Evidence connected to the declared objective and role.
Relate context to declared aims while keeping forecasts and inferred habits distinct.
Explicit objectives, commitments and instructions.
Inferred pattern and explicit correction.
Possible futures are forecasts, not observed facts.
Possible futures remain forecasts rather than observed facts.
Relate proposed work to existing authority and independently observed results.
Possible next action grounded in corrected context.
Actual applicable decision, when supplied; context grants none.
Independent record of what actually occurred under the proposal identity.
Later observed consequence, distinct from the decision and action record.
Relate proposed work to existing authority and independently observed results.
Bounded proposal and pending decision.
Supplied authority, actual action record and independent consequence when present.
The exact proposal, applicable decision, actual action and later consequence retain distinct evidence.
Selected roles, accounts and responsibility sources.
Keep people, roles and delegation sources distinguishable.
Selected people, accounts and roles.
Delegation and responsibility sources, not inferred authority.
A relationship or inferred role does not create authority.
Actual source inclusion and access; not a UI inspection event.
Independent dated observation.
Make inclusion, exclusion, access and freshness explicit.
Explicitly included sources and exclusions.
Dated observation, source and access state.
Opening an explanatory panel does not include a source or grant access.
Source-bearing observation with time and scope.
Selected relevant role and responsibility.
Separately supplied evidence correcting an earlier inference.
Attributed action/consequence history and retained uncertainty.
Retain facts, claims, decisions and corrected interpretations with their lineage.
Evidence, time and unresolved contradiction.
Current inference differs from source observation; correction preserves ancestry.
An inference remains distinguishable from its source observation.
Selected assets, capabilities and dependencies.
Relate selected assets and dependencies to observations and responsible operators.
Asset, dependency and operator relationships.
Known capability, unavailable integration and unknown state.
A known capability or dependency is not an execution grant.
Known capabilities and operator/dependency relationships; no permission.
Explicit objective, commitment or corrected preference.
Evidence connected to the declared objective and role.
Relate context to declared aims while keeping forecasts and inferred habits distinct.
Explicit objectives, commitments and instructions.
Inferred pattern and explicit correction.
Possible futures are forecasts, not observed facts.
Possible futures remain forecasts rather than observed facts.
Possible next action grounded in corrected context.
Actual applicable decision, when supplied; context grants none.
Independent record of what actually occurred under the proposal identity.
Later observed consequence, distinct from the decision and action record.
Relate proposed work to existing authority and independently observed results.
Bounded proposal and pending decision.
Supplied authority, actual action record and independent consequence when present.
The exact proposal, applicable decision, actual action and later consequence retain distinct evidence.
An illustrative context landscape relating sources, responsibilities, capabilities and possible futures.
Additional mechanisms are described in the technical reference.
Your environment connects the sources and capabilities you choose.
People, organizations, roles, accounts and delegated responsibilities.
Inspect the active identity, who may act for whom and the source of an attribute.
Selected documents, repositories, communications, applications, devices and services.
Choose connected sources, exclusions, freshness requirements and access permissions.
Claims, decisions, episodes, contradictions and the intervals when they apply.
Follow the evidence, correct a claim and define retention and deletion behavior.
Assets, services, dependencies, capacity and operating conditions.
See the topology, current observations, unknowns and responsible operators.
Objectives, priorities, obligations, success criteria and resource limits.
Change what the environment is working toward and how success will be judged.
Your explicit preferences and supported patterns in authorized work.
Correct an inference and distinguish it from an instruction you deliberately gave.
Forecasts, horizons, uncertainty and alternative actions.
Examine assumptions and keep a scenario distinguishable from an observation.
Permitted actions, consent, effect identities, results and recovery obligations.
Inspect what may happen, what happened and what remains uncertain; stop or revise the work.
Observations keep their sources. Inferences remain open to correction. You can inspect the model and change what it is allowed to do.
Explore memory & knowledgeThe bitemporal query filters validity time and observation cutoff separately. Source-derived authority is capped through ancestry; equal-authority disagreement remains visible.
The retrieval index is deterministic and rebuildable. Recall activation and consolidation are external-memory operations, not foundation-model weight training.
A correction can update source knowledge, a calibrated estimate, an approved procedure or, in a separately scoped study, an eligible trainable model. These interventions retain different data and evaluation contracts.
The dedicated world model represents selected sources, supported inferences and explicit preferences. It is inspectable and correctable, rather than an unbounded digital copy of a person.
Which date would your compliance lead have to answer for, and could your team reconstruct it today?
Discuss your knowledge environment