canonical-memory

name	canonical-memory
description	Manage canonical (validated, durable) memory in INFYNON Trace. Use when user asks about architecture decisions, API contracts, promotion flows, or auditing canonical notes.
user-invocable	false

Canonical Memory Management — Trace Skill

When to Use

Activate this skill when:

• User asks about "truth memory", "canonical memory", or "validated knowledge"
• User wants to store architecture decisions or API contracts
• User asks what's safe to trust in the codebase
• User wants to promote a note to the highest trust level
• User needs to audit or validate existing canonical notes

What Is Canonical Memory

Canonical memory is the smallest, most durable layer in Trace's three-tier system. It stores knowledge that has been deliberately validated and is intended to be trusted across the team.

Not all memory should be treated equally. "Truth memory which is always correct" is too strong a promise. Even canonical memory can become outdated when code changes. Use "canonical" or "validated" — never promise "always correct."

What Belongs in Canonical

Category	Example
Architecture decisions	"All auth flows go through middleware"
Stable API contracts	"POST /users returns 201 with Location header"
Config invariants	"Rate limit is always per-IP, never global-only"
Known security constraints	"Tokens are RS256 signed, never HS256"
Canonical module facts	"Error handling uses InfynonError enum via thiserror"
Migration rules	"All DB migrations must be backwards-compatible for 2 releases"

What Does NOT Belong in Canonical

• Active caveats or temporary workarounds → team memory
• Personal observations or unfinished thoughts → user memory
• Branch-specific context → team memory with branch scope
• Anything that hasn't been validated across at least one merge
• Anything that changes frequently

Promotion Rules

A note should only enter canonical after:

1. Merge — the related code has landed in main
2. Validation — someone has verified the note matches current code
3. Manual review — a human has approved the promotion
4. Repeated reuse without contradiction — the note has been referenced multiple times without anyone correcting it

Promotion Workflow

Step 1: Identify a candidate

# Find notes tagged for promotion
infynon trace retrieve --tag promote

# Or find long-lived, frequently-referenced team notes
infynon trace retrieve --layer team --tag architecture

Step 2: Validate against current state

Before promoting, verify:

• Is the code still structured this way?
• Has the API contract changed?
• Are there any PRs that contradict this?
• When was the note last updated?

Step 3: Create the canonical note

infynon trace note add canonical-<topic> \
  --title "<Clear, factual statement>" \
  --body "<Details with validation evidence. Include: which PRs validated this, since which version, any constraints.>" \
  --layer canonical \
  --scope repo \
  --tags <relevant-tags>

Step 4: Archive the source note

infynon trace note update <original-id> --status archived

Canonical Note Template

Every canonical note should include in its body:

<Factual statement>

Validated: PRs #X, #Y, #Z
Stable since: vX.Y.Z
Constraints: <any conditions under which this might change>
Last verified: <date>

Auditing Canonical Memory

Canonical notes can go stale. Regularly audit:

# List all canonical notes
infynon trace retrieve --layer canonical

# Check for stale canonical notes
infynon trace retrieve --layer canonical --tag stale

For each canonical note, ask:

1. Is this still true in the current codebase?
2. Has the code it references changed since updated_at?
3. Are there any contradicting team or user notes?

If stale:

# Mark as stale for review
infynon trace note update <id> --status stale

# Or archive if no longer valid
infynon trace note update <id> --status archived

Tips

• Keep canonical memory small — 10-20 notes for a medium project is healthy
• Every canonical note should be something a new team member needs to know on day one
• If you're unsure whether something is canonical, it's probably team memory
• Run infynon trace retrieve --layer canonical at session start to ground yourself
• Use SQL backends for canonical — they're durable, auditable, and queryable