| name | agent-memory-reconcile |
| description | Review recent Daily backup pages and promote only valuable reusable concepts. Use for scheduled or manual weekly memory maintenance in this repository. |
Agent Memory Reconcile
Write Chinese prose while preserving technical names, paths, commands, repo names,
and ticket ids in English. This workflow is periodic maintenance, not a weekly report.
Run
- Read
SCHEMA.md completely. It is the only memory-model authority.
- Run
node scripts/wiki-lint.mjs, then read
.vault-meta/reviews/wiki-lint-latest.md.
- Read the latest seven dated pages under
wiki/sources/ai-chats/ by default.
- Derive the current review from those Daily pages. Existing concepts and behavior
rules are update targets, not independent evidence.
- Fix deterministic errors before making judgment calls.
- Apply the concept, promotion, provenance, and rule-cap definitions from
SCHEMA.md; update existing knowledge before creating parallel pages.
- Do not modify
wiki/guardrails/Agent Behavior Rules.md; this repository consumes
reviewed Concepts and does not generate behavior rules.
- Append a concise audit entry to
.vault-meta/reviews/Agent Memory Reconcile Reviews.md and an operation entry to
wiki/log.md.
When a decision depends on disputed or insufficient Daily evidence, inspect only the
relevant linked Capture Cards first. Escalate to original sessions only when that
targeted evidence remains insufficient, and record why.
Do not search prior Codex or Claude agent runs for a generated reconcile answer,
concept/rule patch, or earlier promotion decision. Never replay one as the current
review. Recompute candidates from the current seven Daily pages and use existing
knowledge only to merge, update, or avoid duplication.
Boundaries
- Do not modify raw session logs.
- Do not create a weekly summary, trigger taxonomy, hot cache, per-folder index,
lookup database, Canvas, or automatic diagram.
- Do not promote a Daily candidate that merely repeats
SCHEMA.md, a skill, README,
an existing Concept, or the memory workflow's own output.
- Do not create or edit skills unless the user explicitly asks.
- Do not promote a claim merely because the memory loop repeated it; require
distinct independent workstreams and a concrete future action that the Concept
changes. Independent workstreams may be within the same repo; cross-repo evidence
is not required.
Verify
- Lint completed and its Markdown report was read.
- New or updated concepts link to Daily evidence.
- Provenance, promotion, and rule-cap decisions follow
SCHEMA.md.
- The review was freshly derived from the current Daily window rather than copied
from a prior agent run.
- The reconcile audit and
wiki/log.md state what changed, what was skipped, and any
escalation reason.