Governed Memory activates only when continuity can improve the task and the requested scope, retention class and write authority permit a specific memory operation. At the 200-turn compaction boundary it commits authorized retained state and content-free source anchors; any later evidence search is delegated to Retrieval.
Governed Memory
Remember what matters — with scope, provenance and deletion.
Separates recent conversation state from reviewer-approved durable facts, with scoped checkpoints, opaque source anchors, correction, retention and explicit deletion.
Approved facts, entities, claims and navigation metadata from the exact allowed scope, with explicit lifecycle decisions and a content-free checkpoint outcome. Retrieval may reopen authorized source evidence, while Memory alone decides what reviewed information becomes durable, corrected, expired or forgotten.
Why it exists
The problem this module is designed to solve.
Stateless assistants repeatedly lose important context, while systems that remember everything create a different risk: incorrect inferences, hidden profiles, cross-conversation leakage and data that cannot be meaningfully corrected or forgotten.
02 / How it works
A bounded path from need to accountable result.
The public model below describes responsibilities and decisions, not sensitive implementation details, provider secrets or customer data.
- 01
Classify the memory
Separate recent conversational state from durable approved user or tenant facts before anything is stored.
- 02
Checkpoint before compaction
Before working context is discarded, atomically commit the retained state and its opaque projection. The checkpoint holds only opaque authorized source references and SHA-256 anchors; source bodies and indexes remain outside Memory.
- 03
Authorize the Retrieval handoff
Expose only the retained anchors allowed for this scope. Retrieval owns source selection and ranking; Memory receives returned evidence only for bounded continuity or an explicit review decision.
- 04
Review then approve a temporal fact
A bounded deterministic queue may propose a direct source relation with exact evidence references; retain it only when a reviewer enters the subject, predicate and value with the selected retention scope and validity interval.
- 05
Correct, expire or forget
Preserve lifecycle actions and provenance so a durable fact can be updated, superseded, expired or removed predictably.
Memory / v1 continuity runtime
Continue safely after compaction — without copying a transcript into memory.
The cognitive runtime now commits the authorized source turns, content-free checkpoint and compacted state as one scoped operation. Concepts needing an agent authority or a customer navigation surface are not shipped as dormant contracts.
4narrow v1 contracts
0raw turn bodies in a checkpoint
64maximum turn anchors per checkpoint
1rule: do not compact when capture fails
What the new layer adds
Continuity that is useful without inventing new authority.
Every record is partitioned by tenant, user and conversation. The contracts fail closed for invalid opaque references, cross-scope reads, conflicting IDs and time regressions.
Original turns stay with their source store
At compaction, the runtime writes original turns only to its scoped source store. The checkpoint receives opaque turn references and SHA-256 anchors, never a body, prompt, summary or source path.
Resume from a safe projection
At the 200-turn history boundary, the compacted state, retained-memory references and immutable turn anchors commit together. A restart verifies the latest same-scope checkpoint before an authorized source reopen.
Never continue from unverifiable context
On restart, Mentaview verifies the scoped checkpoint and source digests before an authorized source reopen. A missing source, collision or mismatch halts the operation.
Deletion includes the continuity attachment
Administrative memory erasure clears checkpoints, source turns and runtime pointers in the same scope. A stale pre-erasure capture cannot be committed afterwards.
Contract state today
Integrated runtime path, bounded product claim.
- The runtime, private local store, PostgreSQL store and fenced server turn journal share the same checkpoint commit.
- Checkpoint writes are idempotent; collisions, invalid anchors, stale timestamps, missing sources and digest mismatch fail closed.
- Administrative memory erasure also removes continuity sources and checkpoint pointers; stale pre-erase batches are rejected.
03 / Customer and operator value
Creates durable continuity without silently mixing conversations, tenants, raw documents, search indexes or inferred preferences.
Exact conversation, user and tenant isolation
Source-backed checkpoints at the real compaction boundary
Review-gated durable memories from fixed markers or explicit declarations
Preview-first memory batches with exact SHA-256 confirmation and all-or-none commit
Review-gated entity candidates and aliases with current evidence
Review-gated temporal and direct-relation claims
Caller-managed wing, room and drawer navigation metadata
Scope-bound memory erasure with retained-evidence checks
Self-contained signed-in memory console with in-memory-only bearer handling
Bounded metadata-only drill-down for managed drawers
Explicit caller-owned tunnels between populated memory rooms
Stable bounded multi-hop traversal with complete tunnel-ID path evidence
Receipt-bound structural topology with stable edge IDs and complete path proofs
Evidence-bound person, project and concept aliases with atomic Unicode-coordinate conflict prevention
Exact alias-aware evidence timelines without inferred identity merges
Usage-driven entity-edge salience with a permanent evidence floor and deterministic decay
Review-only tunnel and hallway proposals from approved entities
Content-free continuity checkpoint availability before explicit wake-up
Reviewable people, projects and concepts placement from approved evidence
Scope-local operational memory and navigation-graph status
Temporal-claim health statistics without source bodies
Explicit temporal claim closure and source-backed successor replacement
Bounded correction lineage from explicit successor links only
Durable reviewed claim links with endpoint evidence and exact scope isolation
04 / Where it creates value
Concrete situations, not generic feature claims.
These are representative product situations. Every deployment still requires its own policy, data boundary and acceptance criteria.
Long-running professional workflows
Resume a project or case with the approved facts and recent conversational state that actually belong to it.
Enterprise assistant preferences
Retain consented user or tenant instructions without turning every past message into an invisible behavioral profile.
Correctable operational knowledge
Carry forward decisions and known facts while preserving a clear path to amend or delete outdated information.
05 / Role in the cognitive system
A clear responsibility creates a trustworthy boundary.
No module is allowed to become an invisible monolith. It owns a narrow contract, composes with named capabilities and refuses responsibilities that belong elsewhere.
What it owns
- Recent conversational state
- Durable reviewer-approved facts
- Reviewer-approved entities and aliases
- Reviewer-approved temporal and direct claims
- Lifecycle of Memory-retained anchors
- Scoped source navigation metadata
- Retention, correction and deletion semantics
- Pre-compaction checkpoint contracts
What it composes with
Cognitive EngineRetrieval as source searchMultimodal Ingestion as a source producerResearch as external transportCoordination as a separate task subsystemPostgreSQL or local stores
What it refuses to own
Boundary before convenience.
Memory does not discover files, parse sessions, admit or delete source records, detect duplicates, build or repair indexes, rank evidence, browse external providers, publish answers or coordinate tasks. Multimodal Ingestion owns source discovery, parsing, admission and removal. Document Retrieval owns source catalogues, indexes, search, ranking, backend reads and index recovery. External Research owns network search and egress. Answer Assurance owns publication checks. Cognitive Engine owns cross-module context composition. Coordination remains a separate subsystem. Memory accepts only already-authorized source anchors and persists only reviewer-approved continuity state and navigation metadata.
06 / Vision and mission
Continuity without hidden accumulation
Memory makes AI more useful over time while supporting the vision of inspectable and sovereign systems: continuity remains scoped, correctable and independent of any model provider.
AI systems that remain useful, inspectable and sovereign across models, providers and deployment boundaries.
Build the cognitive layer that chooses the smallest sufficient path and turns evidence into accountable action.
Capture the value of advanced AI without surrendering data control, architectural freedom or intellectual honesty.
07 / Evidence and maturity
What the current label means — and what remains open.
ENGINEERING MATURITYM4Mentaview internal laboratory assessment
OBJECTIVE REVIEW
Passed with moderate confidence
Scoped memory, conflict handling, restart continuity and authenticated erasure/export paths are locally implemented and exercised; broad adaptive activation is not qualified.
OPEN LIMITS
- No equal-context answer-level representative holdout for broad adaptive activation.
- Production multi-user isolation, resource budgets and canary evidence remain open.
- Customer consent, retention and deletion exercises remain unproven.
What exists today
The Memory crate implements scoped lifecycle, reviewer-gated facts, entities, aliases and temporal claims, content-free checkpoints, correction lineage, caller-managed navigation metadata and erasure across local and PostgreSQL adapters. After Ingestion persistence, the runtime now invokes Memory’s deterministic non-sensitive candidate derivation and returns a content-free handoff with exact Retrieval source roots and separate review and promotion endpoints; deferred recovery republishes it. Memory still never parses the file or promotes a candidate automatically. Focused Memory tests cover restart, isolation, checkpoint verification, exact-source review, replay idempotency, navigation metadata, explicit tunnels, temporal correction, entity evidence and administrative erasure. Ingestion, Retrieval, Research, Answer Assurance, Cognitive Engine and Coordination evidence is displayed and scored under those owning modules rather than used to inflate Governed Memory.
Controlled pilot
A bounded runnable surface exists and can be evaluated in a controlled engagement. Production-scale controls and operating evidence remain gated.
What must earn promotion
Exercise the new post-ingestion review handoff under representative crash recovery and concurrent PostgreSQL operation, then qualify customer consent, retention controls, administrative deletion drills, contradiction, correction, multi-user isolation, production resource budgets and a user-visible resume journey.
Public truth boundary: M4− is a non-standard Mentaview engineering label for internal laboratory validation. It is not an official TRL decision and does not assert production qualification, customer acceptance, independent assurance, certification or universal performance. Inspect the complete assessment record.