13
Schema files
Records, sources, entities, links, policies and telemetry.
Memory architecture · compiled seed workbench
Memory that can be traced, revised and safely used.
A workbench for representing memories as sourced, time-aware records connected to people, projects, concepts and events. It separates what was stated, what was inferred, what changed, and what is safe to retrieve.
Operating chain
Capture→Validate→Connect→Retain→Retrieve→Verify
Seed examples are illustrative. This standalone HTML is not connected to a live memory store.
Showing all 6 searchable sections.
6
Lifecycle stages
Capture through verified use and review.
5
Trust controls
Provenance, conflict, permission, sensitivity and deletion.
Seed
Implementation status
Interactive static prototype; no live persistence.
What counts as a memory?
| Explicit fact | A statement tied to its source and time. |
| Preference | A stated or carefully evidenced stable choice. |
| Episode | An event with participants, time and context. |
| Project state | A requirement, decision, blocker or verified outcome. |
| Concept | A semantic node linked to related concepts and aliases. |
| Inference | A derived hypothesis explicitly labelled as inference. |
| Procedure | A reusable method with scope, preconditions and version. |
Record contract
Every usable memory should carry a stable ID, content, type, subject, source, timestamp, confidence, sensitivity/access policy, validity window and lifecycle state.
contentsourcecontexttimeconfidencepermissions
Missing provenance must remain visible; an unverified record must not quietly become a fact.
Memory lifecycle
Captured→Validated→Linked→Active→Reviewed / superseded→Archived or deleted
| Transition | Trigger | Required evidence | Failure route |
|---|---|---|---|
| Captured → validated | Schema and source checks pass | Source locator, timestamp, required fields | Quarantine incomplete record |
| Validated → linked | Entity and context resolution | Valid IDs and relationship types | Flag orphan / ambiguous link |
| Active → superseded | Newer authoritative statement | New source and explicit revision link | Keep conflict visible |
| Active → archived | Retention or relevance condition | Policy and reason | Keep out of default retrieval |
| Any → deletion workflow | Authorised deletion request | Scope, affected copies, completion receipt | Escalate incomplete propagation |
Schema suite
The downloadable package contains 13 CSV schemas with seed rows and a graph JSON export.
| File | Purpose |
|---|---|
01_MEMORY_RECORDS.csv | Canonical memory records, status, validity and sensitivity. |
02_SOURCES.csv | Source locators, trust class and access policy. |
03_ENTITIES.csv | People, projects, systems and canonical entities. |
04_RELATIONSHIPS.csv | Typed edges with confidence and validity. |
05_CONTEXT_INDEX.csv | Context keys and retrieval hints. |
06_MEMORY_EVENTS.csv | Append-only lifecycle event ledger. |
07_REVISIONS.csv | Supersession and revision history. |
08_RETRIEVAL_RULES.csv | Recall policy and conflict handling. |
09_RETENTION_POLICIES.csv | Retention, review and deletion rules. |
10_GOVERNANCE_RULES.csv | Integrity, privacy and evidence enforcement. |
11_FAILURE_MODES.csv | Orphaned, stale, conflicting or unsafe memories. |
12_OBSERVABILITY.csv | Coverage, precision, staleness and deletion measures. |
13_DEPLOYMENT_MANIFEST.csv | Import order and validation steps. |
Relationship graph
Memorysourced_bySourcesupportsEntityappears_inContext
Relationships should include type, confidence, source and validity dates. Examples: supports, contradicts, supersedes, refers_to, derived_from, applies_to, occurred_during and owned_by.
Time and contradiction
Version Asuperseded_byVersion B
Do not overwrite historical values silently. Preserve what was believed, when it was valid, what changed, why it changed and which source supports the update.
Trust rules
| No source | Mark unverified; limit use. |
| Conflict | Preserve both values and surface disagreement. |
| Low confidence | Qualify, corroborate or ask. |
| Restricted context | Enforce access before retrieval. |
| Sensitive data | Apply strict write, access and retention controls. |
| Deletion | Remove from active retrieval and track governed copies. |
| Generated inference | Label as inference; never present as remembered fact. |
Failure routing
| Orphan | Quarantine and resolve references. |
| Stale memory | Revalidate before use. |
| False recall | Require a matching source record. |
| Permission failure | Deny retrieval and record the event. |
| Incomplete deletion | Keep the deletion workflow open until checked. |
Runtime and deployment
Capture
Ingestion adapters
Conversation, document, event, API and user-authored input connectors.
Memory store
Records + graph
Structured records, append-only event ledger, relationship index and source registry.
Recall
Policy-bound retrieval
Context filtering, semantic search, recency, confidence and provenance checks.
Ingestion → schema validation → provenance check → entity resolution → conflict detection → policy evaluation → indexed memory → retrieval → source-grounded response → feedback/revision
Verification boundary: this deliverable is a static, browser-side prototype. It exports illustrative CSV/JSON files, but does not connect to a database, automatically ingest conversations, persist edits, or enforce real access controls. Those require a backend and tested adapters.