Code Memory 0.9.3

Code Memory 0.9.3

Efficiency update for Codex and Claude Code, plus retrieval fixes that also apply to DeepSeek Harness 0.2.0-rc.2. Unofficial community project, independently maintained and not endorsed by DeepSeek.

What changed

• The installer now adds the auto-context hook for Codex (.codex/config.toml) and Claude Code (.claude/settings.local.json, or .claude/settings.json with -ClaudeScope project). Earlier installers added only the MCP server and skill, which measured no savings in either client. Server-only installs are upgraded in place; remove takes the hook out again.
• Windows hook commands are written for the shell each client uses: cmd /d /c call "…node.exe" … for Codex (PowerShell session or cmd.exe /C), cd . && "…node.exe" … for Claude Code (Git Bash or cmd.exe). Paths use forward slashes.
• Hook guidance: the shown files are byte-exact; make the edit in the first tool call and run the check right after it. Claude Code gets an Edit-tool wording.
• Claude Code context limit: the hook keeps additionalContext under Claude Code's 10,000-character inline limit, above which the client moved it to a file the model had to read first.
• Hook daemon: its address includes the build, so an update never talks to a daemon left from the previous version.
• Retrieval:
  • absolute paths in a request (a check script, a temp directory) are dropped from the query; a random run id in one used to shuffle the ranking;
  • outside a named file, only identifier-shaped names anchor the named route (a minified function A matched the article "A");
  • a short file containing a match is shown whole before dependencies take the budget;
  • requests for a change across uses (wherever, everywhere, every import, across) also get the matched module's users and what they call, with 1.5× the budget;
  • a function in an already found file whose name holds a rarely used query word joins the roots when no chosen root is named after the request;
  • rename requests list every code occurrence of the name (definition, imports, calls, references) from the index.

Measured results

Fresh A/B sessions against the same client without CM, with independent graders that compare each repaired module with the original code:

Project, tasks	Codex (GPT-6 Luna) tokens / time	Claude Code (Sonnet 5.5) tokens / time
Geo Dash, 14 tasks	−58.0% / −35.9%	−40.4% / −35.1%
Geo Dash, 3 larger tasks	−60.1% / −39.2%	−35.0% / −16.4%
node-semver 7.6.3, 6 tasks	−59.1% / −49.0%	−38.2% / −18.0%
superstruct 2.0.2, 6 tasks	−39.2% / −27.4%	−28.6% / −12.0%

Correct solutions were equal or higher with CM in every series. Harness regression check on six Geo Dash tasks: −70.5% tokens, −19.3% time, 11/11. Limits and the weak spots (superstruct in Claude Code, where the price falls only 8.3%; statistics-linear's key-coercion trap) are described in NATIVE-CLIENTS.md. Windows hook commands were checked with Windows Node.js under Wine cmd.exe and with PowerShell 7, not yet in real Windows clients.

Install or update

dsh plugin --profile web add github:kobubu/Rag-Coding#v0.9.3

For the Windows bootstrap, download CM-0.9.3-Windows.zip and SHA256SUMS.txt, verify and extract the ZIP, then run install_cm.ps1 with the same project and -InstallRoot used previously. Rerunning the installer adds the Codex and Claude Code hooks. Codex asks to trust the hook through /hooks. Restart running clients afterward. See the Windows guide.

Node.js 24 or 22.19+ within 22.x is required.