Canonical AI Memory is the reviewed memory an AI can rely on for a bounded project. On LlmWikis.org, this is explained as a practical layer map: preserve raw sources, compile reviewed wiki memory, optionally derive graph projections, export compact AI Memory, and use Project Handoff when responsibility moves. UAIX remains canonical for AI Memory and Project Handoff definitions.
Canonical AI Memory layer map
| Layer | Primary job | Do not confuse it with |
|---|---|---|
| Raw sources | Preserve original reports, source docs, chats, logs, and intake files with path, owner, date, sensitivity, and disposition. | Agent instructions or public truth. |
| Reviewed LLM Wiki | Store durable source memory: compiled pages, source links, trust labels, contradictions, indexes, logs, and long rationale. | Hot operating memory or automatic publication. |
| Derived graph | Project reviewed pages into stable IDs, claim nodes, source spans, review events, contradictions, and retrieval routes. | A hosted graph database, public SPARQL endpoint, or new authority layer. |
| AI Memory | Carry the compact operating packet: current facts, constraints, owners, decisions, next actions, checks, and pointers to deeper sources. | A full institutional archive. |
| Project Handoff | Transfer working context across repository, team, vendor, or agent boundaries. | An agent runtime, scheduler, or permission to skip local rules. |
| Execution agents | Do the work using local tools, project instructions, and human approval boundaries. | Source authority merely because the answer is fluent. |
Build order
- Save source material first. Keep originals or durable source pointers when wording, evidence, or audit history may matter later.
- Compile reviewed wiki pages. Add frontmatter, source links, trust labels, sensitivity, owner, review state, and freshness rules.
- Derive graph projections only after review. Stable IDs, claim nodes, and source spans help retrieval, but they should be generated from reviewed pages and handoff records.
- Export hot AI Memory carefully. Include only current facts, constraints, decisions, owners, next actions, checks, and links to deeper sources.
- Use Project Handoff for transfer. When work changes hands, include the front door, typed memory records, intake state, constraints, expected touchpoints, and targeted checks.
- Write accepted outcomes back through review. A useful agent result becomes durable memory only after source, privacy, authority, and verification checks.
Review rule
A report, dropped file, old chat, generated summary, wiki page, or graph edge is not canonical by default. It becomes Canonical AI Memory only when the project promotes an accepted slice into current AI Memory, Project Handoff, durable docs, code, tests, release notes, roadmap state, machine artifacts, or long-memory evidence.
Current support boundary
LlmWikis.org explains this pattern; UAIX defines the canonical AI Memory and Project Handoff surfaces. This page does not claim hosted .uai upload validation, automatic repository writes, automatic LLM Wiki sync, hosted graph services, public graph APIs, public SPARQL endpoints, SDKs, CLIs, official adapters, certification, endorsement, or UAI-1 conformance.
When to use each layer
| Need | Use | Why |
|---|---|---|
| Original wording, provenance, or future audit. | Raw source archive. | Summaries can lose the exact evidence. |
| Deep source-backed knowledge and long rationale. | Reviewed LLM Wiki. | It can hold broad context without bloating hot memory. |
| Navigation across claims, sources, contradictions, and review events. | Derived graph or graph-aware index. | It improves retrieval while preserving source citations. |
| Next agent or maintainer needs current operating truth. | AI Memory packet. | It is compact enough to load before work. |
| Responsibility, execution, or ownership is moving. | Project Handoff. | It includes transfer-specific constraints, touchpoints, and checks. |
Setup wizard fit
The setup wizard should capture the Canonical AI Memory choices that affect future readers: raw source path, compiled wiki path, index/log path, source-routing strategy, graph and retrieval boundary, promotion owner, evidence destination, Project Handoff use, workspace routing, and blocked actions. The output remains browser-only planning data until a human applies it to real files.