Share project memory between Codex and Claude Code
Use one reference for project decisions and current state without copying entire conversation histories.
Last reviewed:
About this guide
This guide was prepared with AI assistance using the project’s public documentation. It does not claim testing on every device or independent human review. Project publisher: AmirSina Zamanian.
Sources are listed at the end. Use the project repository to report an error or suggest a correction. Report an issue
⌕
Search runs on this device.
On this page
What does shared memory mean?
Shared memory means both tools consult the same verified project state; it does not merge their internal model memory or chat history. In Personal Agentic OS, tool instruction files direct the agent toward the project context pack. That pack should retain the goal, decisions, open questions, and next action with provenance. Creating two instruction files alone does not demonstrate that either model has read them.
Choose a non-sensitive project and a separate memory destination. Select only codex and claude during setup and review the proposed project map. When both candidates point to the same project path, inspect that they resolve to one canonical project. If paths differ, do not rely on similar names alone. In our published synthetic experiment, two candidates with the same path resolved to one project.
Start with a non-sensitive project. In Claude Code, record one accepted decision, its source file, one unresolved question, and the next action in the downloadable handoff template. Mark unrun tests as not run. Save and inspect the file yourself. Open the same project in Codex and give it the exact approved file path. Ask: Read this handoff without editing files. Restate the goal, accepted decision and source, unresolved question, and next action. Tell me what is missing; do not infer permission to execute. Compare its answer with the file. A wrong decision, missing source, or invented test result means the handoff failed: correct the record before allowing edits. You can reverse the tool order. This manual exercise needs no PAO installation and is a procedure to try, not a claim that we have tested live model sessions. With PAO, perform approved recording and sync, then apply the same read-back check to the generated project context pack.
In 0.12, use bootstrap /absolute/path/to/config.json <project-id> for a short start-of-session summary. It prioritizes confirmed decisions and reports how many records it omitted; it is not a substitute for the complete pack. If a needed decision or source is absent, page through Decision with details or inspect that project's full pack with full-context. Ask the receiving tool to repeat the project ID and decision source; adapter creation alone does not prove Codex or Claude Code read it.
This workflow is not a replacement for backups or access controls. A local file does not imply a local model: content supplied to a cloud-connected tool may be sent under that provider’s configuration. We claim neither token savings nor higher accuracy without a matched comparison. Use the installation guide and reproducible lab below for a safe first check.