agentsclimarketplace

Codex project memory

Skill Retourflow/project-memory-for-codex/skill/codex-project-memory

Verified, token-aware project memory for Codex with selective loading, context budgets, and Git-reviewable Markdown.

Install
npx -y skills add Retourflow/project-memory-for-codex --skill codex-project-memory

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Maintain a minimal, token-aware, verified project memory for Codex repositories. Use when initializing project context files, resuming work across sessions, changing architecture or project goals, recording decisions, updating current state after implementation, or reducing bloated and stale agent documentation.

SKILL.md

3.7 KB, as published. Nobody here has run it

Codex Project Memory

Preserve the smallest verified project state needed for reliable future work.

Core Rules

  1. Treat code and tests as evidence; do not turn assumptions into project memory.
  2. Keep always-loaded context small.
  3. Read detailed documents only when the task needs them.
  4. Replace stale state instead of appending an endless activity log.
  5. Mark future work as Planned and shortcuts as Temporary.
  6. Use Git history for exhaustive history; do not duplicate it in active context.
  7. Never store secrets, credentials, tokens, personal data, or private machine-specific information in project memory.
  8. Do not bypass sandboxing, approval requirements, repository permissions, or other user-controlled security boundaries.

Start A Task

  1. Read the nearest applicable AGENTS.md.
  2. Read docs/PROJECT_CONTEXT.md and docs/CURRENT_STATE.md.
  3. Inspect relevant code before relying on documentation.
  4. Load additional context by task type:
    • Architecture, interfaces, data flow, storage, dependencies: read docs/ARCHITECTURE.md.
    • A choice whose rationale matters: read relevant entries in docs/DECISIONS.md.
    • Historical investigation only: search targeted entries in docs/CHANGELOG.md or Git.
  5. If documentation conflicts with implementation, verify behavior and repair the documentation as part of the task.

Finish A Task

Update only documents affected by verified changes:

  • Goal, scope, audience, constraints: PROJECT_CONTEXT.md
  • Current progress, active work, blockers, next steps: CURRENT_STATE.md
  • Implemented structure, interfaces, data flow, dependencies: ARCHITECTURE.md
  • Significant choices and trade-offs: DECISIONS.md
  • Concise user-visible or architectural change summary: CHANGELOG.md

Then:

  1. Remove or replace stale statements.
  2. Keep each document within the budgets defined in AGENTS.md.
  3. Review the memory diff for sensitive or proprietary information.
  4. Run python scripts/validate.py <project-root> when the script is installed.
  5. Report which memory documents changed and why.

Decision Test

Create a decision record only when reversing the choice later would be costly or confusing. Include context, decision, rationale, alternatives, consequences, date, and status.

Do not create decision records for routine edits, formatting, or obvious implementation details.

Compaction

When a document exceeds its budget:

  1. Preserve current truth, non-obvious constraints, and active risks.
  2. Remove completed task narration from CURRENT_STATE.md.
  3. Collapse superseded architecture descriptions into the current design.
  4. Move durable rationale into DECISIONS.md.
  5. Leave detailed historical reconstruction to Git.

Evidence And Claims

When evaluating whether project memory saves context:

  1. Prefer reproducible measurements over intuition.
  2. Label character-based token calculations as estimates.
  3. Keep static fact coverage separate from agent task success.
  4. Report real token usage or success rates only when supplied by an actual run.
  5. Never infer model quality from context size alone.

If the repository includes benchmarks/run.py, run it after changing document routing, budgets, benchmark tasks, or report methodology.

Read references/document-contracts.md when initializing files or deciding where information belongs.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.