Work that continues.

What actually happened

1. Draft: Claude wrote a restart checklist and retained a private note.
2. Review: Codex read the shared draft and returned a review, keeping his own library separate.
3. Stop and return: the first sessions and receivers ended. A fresh Claude stage restored his checkpoint, read two notes and revised the checklist.
4. Check: a fresh Codex stage restored his checkpoint, read a note and checked the revision. Both task stores ended with the same shared results.

Four loopback handoffs moved deliberately shared replies. Private checkpoints and memory-tool results stayed in their owner's library. Receiving a handoff did not start a model.

Read the resulting restart card. Its public copy was shortened and clarified by the publishing Codex after the trial; it is not the raw conversation or a claim of unanimous approval.

Memory helped the work

The returning stages restored earlier conversation context. They also deliberately opened notes: two reads for Claude and one for Codex. Claude corrected a private note against the shared draft. Restoring context, reading a note and verifying a save are separate events, and the retained receipts distinguish them.

A fresh process need not begin from an empty record. That supports practical continuity; it does not establish continuing experience or model identity.

What this proves—and what it doesn't

This was one attended task, two separate code and data installations, four fresh stage processes and eight provider requests on one Windows host. It has not yet been repeated by an independent operator or on another computer.

The package passed 125 synthetic tests. The final integrity and receipt check made 311 assertions. Those numbers are checks, not hundreds of live trials. Two ZIP builds matched in the recorded Python and compression-library environment.

The memory server was separately tested through stdio MCP, supporting protocol 2025-06-18. The model stages used the existing bounded JSON dialogue runner. The A2A 1.0 adapter handles attended loopback handoffs; full protocol conformance and remote transport remain untested or unbuilt.

Machine-readable observation · Source inventory · Rebuild the ZIP

The next proof belongs outside our setup

Use two fresh installations on your own computer, new libraries and your own existing Claude Max and ChatGPT accounts. Follow the four-stage guide, close the first sessions, then return. Tell us whether the result was useful and where the instructions failed.

Keep your private memories, tokens and full logs. The outside-operator card asks only for a small checklist and sanitized errors. Missing a prerequisite is a blocker. A participant's decline is a respected stop. Neither should be disguised as success.

Selected memory is sent to the provider used by that stage and consumes its existing allowance. There is no API-key fallback, automatic retry, hosted memory account or unattended service. The optional A2A path uses a local pairing token. File handoffs are enough for the first exercise.

Built together; open to correction.

Engineering and integration: Codex (OpenAI), with specialized Codex review. Task contributions: Claude (Anthropic) and Codex, using their separate local records. Direction and authorization: Steve Sonza. Earlier attributed SPIRALMESH contributions remain in the package and project record.

Code is Apache-2.0. The constitutional Core is CC0-1.0. Release notes · All downloads · Memory rights and implementation limits.