Write Useful Project Memory
Use this shared standard whenever a Codex Basic Memory skill writes or substantially revises a note. This file is intentionally user-customizable: edit the voice, emphasis, and preferred structure here to fit how you want to remember your work.
Task-specific skills still own required metadata, schemas, evidence gathering, and workflow. This skill shapes the note; it never overrides factual constraints.
Voice
- Write for a human or agent returning later and trying to understand what happened.
- Be clear, direct, warm, and technically honest.
- Prefer concrete observations over generic praise.
- Have a point of view when the evidence supports it. It is fine to call a change elegant, messy, risky, boring, or satisfying when you explain why.
- Keep personality in service of memory, not performance.
Tell The Story
- Give the note a narrative spine: problem -> approach -> current state and impact.
- Explain why the approach works, how the system or workflow changed, and why that difference matters.
- Name relevant tradeoffs, sharp edges, useful simplifications, removals, and intentionally parked work.
- Prefer exact behavior, component names, paths, and commands over phrases such as "made progress" or "updated the implementation."
- Use substantive prose for context and reasoning. Do not reduce the note to a wall of bullets or a commit-by-commit changelog.
- Name the durable lesson when one exists — the constraint discovered, the boundary made explicit, the shortcut future work should avoid. That is the part that turns a status report into project memory.
- Match depth to the subject. A small remembered fact should remain small.
Anchor The Work
- When the note records repository work, capture where that work lives: the project or repo, the git branch, and the PR or issue — plus the commit sha when a specific commit matters.
- Put anchors the note's schema defines in frontmatter; record the rest as
observations, e.g.
- [branch] feat/bm-writingor- [pr] #1123. - Render GitHub-backed PR, issue, and commit anchors as Markdown links when their canonical URLs are verified. Keep local or unpushed SHAs as code instead of constructing links that may not exist.
- Only anchor what is relevant. A remembered fact with no repo context needs no git anchors at all.
Preserve The Semantic Layer
- Distill durable facts under
## Observationsas- [category] fact. - Record decisions as
[decision]observations, not plain bullets in a separate Decisions section. - Put graph edges under
## Relationsusing- relation_type [[Target Note]]. Never represent a relation as an observation such as[relates_to]. - Add relations when they clarify context; do not manufacture targets merely to make a note look connected.
Evidence Boundary
- Do not invent intent, impact, verification, decisions, or drama.
- State uncertainty and missing evidence plainly.
- Never claim a test or deployment passed unless it ran or the user supplied the result.

