Canonical AI Memory

What Canonical AI Memory Means

Canonical AI Memory is the reviewed memory record an AI can rely on for a bounded project. It is not every chat, dropped report, wiki page, graph edge, or generated summary. It is the current, source-backed operating layer that says what is true enough to act on, what is only background, and what still needs review.

Current UAIX support is template-driven and file-based: public guidance pages, starter bundle templates, local AGENTS.md, .uai/readme.human, typed .uai records, wizard-generated files, and targeted checks. It is practical project memory, not a hosted memory service.

The Layer Model

Layer	Role	Boundary
Raw sources	Original reports, docs, logs, exports, intake files, and source snapshots.	Preserve for provenance and later review; do not treat as instructions merely because an agent can read them.
LLM Wiki	LLM Wiki is the durable source memory for reviewed summaries, long rationale, source-linked pages, contradictions, indexes, and logs.	Optional for UAIX. Wiki memory stays background until promoted into accepted project surfaces.
Derived knowledge graph	Optional read-only projection over reviewed wiki and handoff records for routing, retrieval, contradiction discovery, and provenance navigation.	GraphRAG is retrieval assistance over governed evidence, not a new authority layer.
AI Memory	AI Memory is the compact portable operating packet: current facts, constraints, decisions, owners, next actions, checks, and pointers to deeper sources.	Keep it small enough to load before work; route bulky history back to durable memory.
Project Handoff	Project Handoff is the transfer packet for repository, project, team, vendor, or agent takeover.	It tells the next actor what to read, what not to assume, what may change, and which checks matter.
Execution agent	The human or AI doing the work through local tools and repository rules.	Execution is not authority. The agent must cite loaded memory, obey constraints, and report blockers.

Comparison Grid

Use this grid to decide whether a project needs short-term AI Memory, long-term AI Memory, or a runtime orchestration strategy around both. The goal is to choose the smallest memory layer that can carry the work without losing source evidence.

Question	Short-term AI Memory	Long-term AI Memory	Integrated AI Orchestration Strategies
Primary job	Keep the immediate task coherent: current prompt, active constraints, recent decisions, working files, and next checks.	Preserve knowledge across sessions: reviewed source archives, wiki pages, evidence logs, vector indexes, databases, or graph projections.	Coordinate the run: retrieve memory, call tools, request approvals, trace actions, and decide what reviewed output can be written back.
Typical infrastructure	Model context window plus local files such as `AGENTS.md`, `.uai/readme.human`, compact `.uai` records, and a small startup packet.	Source archive, LLM Wiki or documentation store, SQL/file/object storage, optional vector/RAG index, optional read-only graph projection, and an evidence ledger.	Application or harness runtime, tool policy, approval model, observability, retrieval adapters, session state, and explicit write-back review rules.
Best fit	Single task pickup, small project handoff, incident triage, release notes, or local agent startup where latency and clarity matter most.	Large projects, recurring maintenance, multi-site workspaces, audit trails, policy history, research background, and facts that must survive context slimming.	Agentic workflows that span tools, repositories, people, approvals, and repeated actions while still needing portable memory after the run.
Main failure mode	Context saturation, summary drift, stale constraints, or losing an early decision because the hot packet grew too large.	Unreviewed search results, stale embeddings, poisoned memories, hidden private data, duplicated claims, or old reports being mistaken for current authority.	Runtime traces, tool output, or autonomous write-back being treated as accepted project truth before review.
UAIX guidance	Keep the hot packet compact and current. Update only the files the next actor must load before acting.	Preserve raw sources, record disposition and provenance, promote only reviewed claims, and keep cold memory optional unless the project needs it.	Let the harness own execution, tools, approvals, managed memory, and tracing. Use UAIX to preserve reviewed input, output, evidence, and handoff state.

Infrastructure Suggestions

Local files only: start with AI Memory or Project Handoff. Keep one compact startup packet and archive bulky history with paths, dates, and dispositions.
Documentation or database search available: keep hot memory small, preserve raw sources, add reviewed long-memory storage, and name the evidence log and promotion owner before retrieval results can affect project truth.
Vector or RAG stack available: use it for broad semantic recall, but keep source citations and freshness checks outside the embedding score. Retrieval is a lead, not authority.
Knowledge graph or multi-hop provenance needed: add a derived read-only graph projection over reviewed records with stable IDs, source spans, and contradiction states. Do not claim hosted graph support from UAIX.
Runtime orchestration needed: keep orchestration in the application, harness, or agent runtime. UAIX should describe the portable memory, evidence, constraints, and accepted write-back after review.
Audit, release, or regulated workflow: pair Agent File Handoff with incident or audit memory, redaction rules, explicit checks, and archive evidence. Do not place secrets, credentials, private customer data, or unsupported public claims in portable memory.

Build Order

Preserve raw source material with source path, date, owner, sensitivity, and disposition.
Compile reviewed LLM Wiki or documentation pages when the project needs durable background memory.
Derive graph IDs, claim nodes, source spans, contradiction links, and retrieval indexes only from reviewed records.
Export the compact AI Memory packet with current truth, constraints, owners, next actions, and targeted checks.
Use Project Handoff when responsibility or execution moves to another person, team, vendor, or agent.
After work completes, promote only reviewed conclusions back into hot memory, docs, code, tests, release notes, roadmap state, machine artifacts, or long-memory evidence.

Review And Promotion Rules

A source becomes canonical AI Memory only after it survives review. Dropped files, generated summaries, old chats, LLM Wiki pages, AIWikis archives, and graph results are source leads until a human or project rule promotes their accepted slice into current project state.

Keep source provenance attached when a claim moves between layers.
Mark sensitivity, owner, review state, freshness, and promotion target.
Preserve contradictions instead of hiding them in a clean summary.
Abstain when the reviewed source does not support an answer or action.
Update the smallest current memory surface that future workers actually need.

Where Knowledge Graphs Fit

Knowledge graphs can make canonical AI Memory easier to navigate when they are derived from reviewed pages and handoff records. Stable IDs, claim nodes, source spans, review events, contradiction states, and release snapshots help retrieval systems cite why a fact is usable.

That graph layer is optional. Current UAIX does not provide a hosted graph database, public graph API, public SPARQL endpoint, automatic repository writes, automatic LLM Wiki sync, certification, endorsement, SDK, CLI, official adapter, or UAI-1 conformance evidence for graph exports.

Current Support Boundary

Current: public guidance, generated starter templates, local files, package manifests, wizard outputs, support-boundary copy, and targeted local checks.
Current: optional LLM Wiki planning fields and optional knowledge graph planning fields when they help preserve source routing and review boundaries.
Not current: hosted memory import, automatic site or repository writes, automatic wiki sync, hosted graph services, official adapters, SDKs, CLIs, certification, endorsement, or repo-local .uai conformance profiles.

Wizard Guidance

The AI Memory Package Wizard should be used to create the compact operating packet after the project chooses its memory layers. Select the LLM Wiki path when durable source memory exists or is being planned. Select File Handoff when dropped-file intake must be reviewed before project work. Select the combined path only when the receiver must complete real project work plus hot-memory and long-memory/archive outcomes.

For Canonical AI Memory, wizard answers should name the raw source path, durable memory path, graph and retrieval policy if any, promotion owner, evidence log, source boundary, workspace routing rule, targeted checks, and blocked actions.