# Shared Memory Intake

> Curate shared Codex memory when Claude submits findings or Codex reviews, promotes, or maintains task memory.

- Skill: `ohyeh/shared-memory-intake` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add ohyeh/shared-memory-intake`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ohyeh/shared-memory-intake/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: ohyeh (https://skillmd.com/u/ohyeh)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/ohyeh/shared-memory-intake

---


# Shared Memory Intake

Use `~/.codex/memories/` as the canonical shared knowledge store and
`~/.agents/shared-memory-inbox/` as the external submission store. Read
`MEMORY.md` before opening matching `rollout_summaries/` files. Treat dynamic claims
as `UNCONFIRMED` until verified in the active session.

## Actor contract

Require an explicit actor: `--as claude` or `--as codex`. Never infer it from the
runtime.

| Actor | Allowed operations | Forbidden operations |
|---|---|---|
| `claude` | read official memory; create one submission in `inbox/pending/` | edit or delete official memory, or another submission |
| `codex` | read, review, promote, supersede, maintain | delete historical summaries or silently overwrite a current claim |

## Claude: submit

1. Read `MEMORY.md`; do not restate an existing current claim as a new submission.
2. Create one uniquely named Markdown file under `~/.agents/shared-memory-inbox/pending/` with YAML frontmatter:

   ```yaml
   schema_version: 1
   source_runtime: claude
   submitted_at: 2026-07-26T00:00:00Z
   cwd: /absolute/path
   source_session: opaque-session-id
   claim_status: unverified
   evidence_host: hostname-or-fleet-id
   ```

   `evidence_host` is optional for same-host evidence and required when the cited
   source exists only on another host.

3. Keep only reusable conclusions, source references, evidence, and unknowns. Never
   include credentials, tokens, private keys, cookies, or raw transcripts.
4. Validate before handoff:

   ```sh
   python3 ~/.agents/skills/shared-memory-intake/scripts/validate_submission.py /absolute/path/to/submission.md
   ```

5. A correction is a new submission with `supersedes_submission`; never overwrite the
   earlier proposal.

## Codex: review and promote

1. Acquire the single-curator lease before promotion: `mkdir ~/.agents/shared-memory-inbox/.curator.lock`.
   If it already exists, stop and report the active-curator conflict. Release it with
   `rmdir` after the outcome is recorded.
2. Run the validator on the candidate. Reject on schema or secret-pattern failure.
3. Inspect the cited source and live-verify every dynamic claim that would enter the
   current index. For cross-host evidence, require `evidence_host`, verify the live
   source on that host, and record the curator host plus verification time in the
   promoted summary. Missing host access remains `UNCONFIRMED`.
4. Re-read `MEMORY.md` immediately before its update. If its relevant topic changed
   during review, restart the review instead of merging assumptions.
5. Promote accepted knowledge to a new immutable `rollout_summaries/*.md` record and
   update `MEMORY.md` as the only current index. Give every changed reusable claim a
   lifecycle state: `active`, `superseded`, or `UNCONFIRMED`.
6. Move the candidate to `processed/` or `rejected/` without deleting its audit trail.

## Codex: maintain

1. Acquire the same single-curator lease and re-read `MEMORY.md` immediately before
   every official-memory write; stop on an active-curator conflict.
2. Scan current index entries for changed paths, versions, models, deployments, or
   stale references.
3. Demote unverified dynamic claims to `UNCONFIRMED`; mark replaced claims
   `superseded` with their replacement reference.
4. Consolidate duplicate current entries. Do not rewrite historical rollout summaries.
5. Re-run the validator on any newly promoted submission and report the paths changed.

## Validator

Use `scripts/validate_submission.py`. It is a narrow preflight gate, not proof that a
claim is true and not a complete secret scanner. A pass authorizes review only; Codex
still owns evidence verification and promotion.