# Memory Schema and Noise Budget

One memory bead = one atomic reusable rule.
If it is not reusable across future tasks, keep it in task notes instead.

Canonical memory schema:
- type: decision
- status: closed (after ratification)
- labels: memory,mem-canonical,mem-hard|mem-soft,<domain>
- title format: [MEMORY][<DOMAIN>][HARD|SOFT] <rule sentence>

Required description sections:
- Scope:
- Out of Scope:
- Rule:
- Rationale:
- Failure Mode:

Required acceptance style:
- Given / When / Then invariant
- Verification commands

Required metadata JSON keys:
- memory_version
- memory_strength
- domain
- effective_date
- owner
- supersedes
- superseded_by
- evidence_ids
- plan_refs

Edge rules:
1) Anchor <-> Canonical memory
- Use: bd dep relate
- Never use blocker edge for anchor indexing

2) Active task <-> Memory
- Default: bd dep relate

3) Active task <- Contract (hard rules only)
- Use: bd dep add <active-task> <contract>

4) Memory versioning
- Use: bd supersede <old> --with <new>

Noise budget limits:
- Per active task: 3-7 related memory nodes
- Per active task: 0-2 blocker contracts
- Per canonical memory: 1 primary anchor domain
- Per canonical memory: 2-5 source-bead provenance links
- Promote incident/workaround memory only after repetition evidence

Promotion policy:
1) Incident repeats 2+ times -> candidate soft canonical memory
2) Workaround survives release window -> candidate hard/soft canonical memory
3) Obsolete canonical memory -> supersede, do not rewrite history