Codex Agentmemory Windows

Codex Agentmemory Windows

Goal

Set up agentmemory for Codex App on Windows with PowerShell-safe hook commands, Codex-native memory injection, stable data paths, working Stop finalizers, healthy knowledge graph updates, and repeatable validation. Prefer the bundled scripts for deterministic repair.

Quick Repair

Run the helper when the user wants an automatic fix:

python scripts/configure_agentmemory_windows.py --codex-home "$env:USERPROFILE\.codex"

Use these options when needed:

python scripts/configure_agentmemory_windows.py --secret ""
python scripts/configure_agentmemory_windows.py --url "http://localhost:3111" --secret ""
python scripts/configure_agentmemory_windows.py --preserve-other-hooks
python scripts/configure_agentmemory_windows.py --absolute-data-dir "$env:USERPROFILE\data"
python scripts/configure_agentmemory_windows.py --patch-codex-scratch-context
python scripts/configure_agentmemory_windows.py --skip-injection-patch
python scripts/configure_agentmemory_windows.py --dry-run

The script backs up changed files, creates agentmemory-hook.cmd, writes Windows-safe hook commands, updates Codex hook trust hashes, configures the agentmemory MCP section, patches hook scripts to emit Codex additionalContext JSON, filters Codex Desktop background suggestion sessions, optionally pins agentmemory's data store to an absolute path, and validates the hook commands.

After repair, rerun with --dry-run; it should print only environment/path information and no would update: lines. If hooks.json was edited manually, treat any trusted-hash drift as a configuration repair, not a one-line hash tweak: rerun the helper so hooks.json, config.toml, wrapper paths, and patched hook scripts stay aligned.

Knowledge Graph Repair

Use this when the viewer graph is tiny, stale, or frozen even though observations/summaries keep growing.

First inspect without changing files:

python scripts/repair_agentmemory_graph_windows.py --dry-run

Then apply the full repair:

python scripts/repair_agentmemory_graph_windows.py --all --restart-service

The graph repair script:

• patches every discovered stop.mjs so Codex Stop finalizes the session, checks server-side GRAPH_EXTRACTION_ENABLED, and posts graph/extract in small batches;
• patches installed dist\index.mjs so graph-extract remaps edges from temporary extracted node IDs to persisted node IDs after node deduplication;
• backs up mem%3Agraph%3Anodes.bin and mem%3Agraph%3Aedges.bin before writing backfill data;
• backfills a deterministic graph from existing compressed observations using file/concept/title metadata instead of expensive full LLM extraction;
• validates via /agentmemory/graph/stats after import.

Default backfill settings are intentionally conservative: --min-importance 5 --max-files 2 --max-concepts 3. Raise or lower them only when the user explicitly wants a denser or smaller graph.

Manual Workflow

1. Inspect %USERPROFILE%\.codex\hooks.json, %USERPROFILE%\.codex\config.toml, and %USERPROFILE%\.codex\agentmemory-hook.cmd.
2. Locate the installed agentmemory plugin root, usually %USERPROFILE%\.codex\plugins\cache\agentmemory\agentmemory\.
3. Locate node.exe; prefer an absolute path from Get-Command node.exe, C:\Program Files\nodejs\node.exe, or Codex bundled node.exe.
4. Read AGENTMEMORY_SECRET and AGENTMEMORY_URL from existing config or environment. Do not hardcode another user's secret.
5. Create a wrapper at %USERPROFILE%\.codex\agentmemory-hook.cmd:

@echo off
set "AGENTMEMORY_URL=http://localhost:3111"
set "AGENTMEMORY_SECRET="
set "AGENTMEMORY_INJECT_CONTEXT=true"
if "%~1"=="" exit /b 2
"C:\Path\To\node.exe" "C:\Path\To\agentmemory\scripts\%~1"
exit /b %ERRORLEVEL%

1. Write hook commands through cmd.exe, not direct quoted PowerShell invocation:

cmd.exe /d /s /c ""C:\Users\\.codex\agentmemory-hook.cmd" session-start.mjs"

1. Use this event-to-script map:

1. Update config.toml [hooks.state.'...\hooks.json::0:0'] entries with enabled = true and trusted_hash = "sha256:".
2. Disable duplicated plugin-provided hooks under agentmemory@agentmemory:hooks/hooks.codex.json if the global hooks.json wrapper is used.
3. Patch session-start.mjs, prompt-submit.mjs, and pre-tool-use.mjs so any injected memory is returned as Codex hook JSON:

{
  "hookSpecificOutput": {
    "hookEventName": "SessionStart",
    "additionalContext": "..."
  },
  "suppressOutput": true
}

Use hookEventName: "UserPromptSubmit" for prompt-submit.mjs and hookEventName: "PreToolUse" for pre-tool-use.mjs. Raw `` text on stdout is not enough for reliable Codex App injection. 11. If starting agentmemory from different working directories can create multiple ./data stores, edit each installed dist\iii-config.yaml so state_store.db and stream_store use absolute paths such as %USERPROFILE%\data\state_store.db and %USERPROFILE%\data\stream_store. 12. For graph updates, patch stop.mjs to query /agentmemory/config/flags instead of trusting Codex process env vars, then call /agentmemory/graph/extract in batches after /agentmemory/session/end. 13. Patch service-side graph-extract if needed so merged nodes update edge source/target IDs before edge persistence. 14. Validate by piping sample hook JSON into every command through PowerShell and requiring exit code 0.

Missing Injection Checklist

Use this when capture and summaries work but new conversations do not receive memory:

1. Confirm agentmemory status shows nonzero observations or memories and AGENTMEMORY_INJECT_CONTEXT is enabled.
2. Call the context endpoint or session-start.mjs manually. If it returns a non-empty ``, agentmemory has data and the remaining issue is Codex hook delivery.
3. Inspect the newest %USERPROFILE%\.codex\sessions\YYYY\MM\DD\*.jsonl. A successful injection appears as a developer message containing ``.
4. If manual hook output is raw XML/text but the JSONL has no developer context message, patch the hook scripts to use hookSpecificOutput.additionalContext.
5. If a new Codex scratch folder like %USERPROFILE%\Documents\Codex\YYYY-MM-DD\hi-3 gets empty context while hi-2 has memory, patch the agentmemory service bundle so mem::context matches the daily Codex scratch root instead of exact session.project === data.project.
6. If agentmemory status suddenly shows 0 observations or memories after restart, check for accidental data stores under the current workspace or package directory and pin iii-config.yaml to an absolute data directory.

Duplicate Session Checklist

Use this when Codex App shows one visible conversation but agentmemory shows many sessions with the same title:

1. Inspect recent sessions. If repeated firstPrompt starts with # Overview Generate 0 to 3 hyperpersonalized suggestions..., these are Codex Desktop background recommendation jobs, not duplicate user chats.
2. Treat this as capture noise: the hook is recording every Codex session-like run, including hidden background jobs.
3. Patch prompt-submit.mjs to detect that exact prompt prefix, record the session id in a temp ignore file, call /agentmemory/forget for the pre-created background session, and return without observing.
4. Patch pre-tool-use.mjs, post-tool-use.mjs, stop.mjs, and pre-compact.mjs to skip any session id in that temp ignore file.
5. Add a service-side fallback in installed @agentmemory/agentmemory/dist/src-*.mjs and dist/index.mjs: mem::observe should ignore this prompt pattern and delete the pre-created session; /agentmemory/session/start should also ignore the pattern if it arrives as a title.
6. Restart agentmemory after service-bundle changes.
7. Do not delete old repeated sessions automatically. Ask the user first, then use /agentmemory/forget for sessions whose prompt clearly matches the background suggestion pattern.

Codex Scratch Project Matching

agentmemory v0.9.x may match session context by exact project path. Codex App often creates per-chat scratch folders under:

%USERPROFILE%\Documents\Codex\YYYY-MM-DD\

For shared daily scratch context, patch mem::context in the installed agentmemory service bundle (dist\src-*.mjs and dist\index.mjs) to:

• normalize Windows paths case-insensitively and slash-insensitively;
• treat the same %USERPROFILE%\Documents\Codex\YYYY-MM-DD parent as one context root;
• include semantic memories only when their fact text contains the current project leaf hint, to avoid injecting unrelated project facts.

The helper applies this with --patch-codex-scratch-context when the installed bundle still has the original exact-project matcher.

Why This Fixes Windows Hook Errors

PowerShell treats "C:\path\agentmemory-hook.cmd" as a string. If arguments follow it directly, PowerShell can fail with Unexpected token 'session-start.mjs'. The stable Windows form is either:

& "C:\path\agentmemory-hook.cmd" session-start.mjs

or, more robust for Codex hook runners:

cmd.exe /d /s /c ""C:\path\agentmemory-hook.cmd" session-start.mjs"

Use the cmd.exe form in hooks.json.

Validation Commands

Use this after manual edits:

$payload = '{"session_id":"test","cwd":"C:\\Temp","prompt":"test"}'
$hooks = Get-Content "$env:USERPROFILE\.codex\hooks.json" -Raw | ConvertFrom-Json
foreach ($event in $hooks.hooks.PSObject.Properties.Name) {
  $cmd = $hooks.hooks.$event[0].hooks[0].command
  $payload | powershell.exe -NoProfile -Command $cmd
  Write-Host "$event exit=$LASTEXITCODE"
}

All events should return exit=0.

For injection-specific verification, parse hook stdout as JSON and check:

$payload = '{"session_id":"test","cwd":"C:\\Temp","hook_event_name":"SessionStart"}'
$out = $payload | & "$env:USERPROFILE\.codex\agentmemory-hook.cmd" session-start.mjs
$json = $out | ConvertFrom-Json
$json.hookSpecificOutput.additionalContext

If this contains <agentmemory-context, open a fresh Codex conversation and confirm the newest session JSONL has a developer message containing the same tag.

For graph-specific verification:

Invoke-WebRequest -Uri "http://localhost:3111/agentmemory/graph/stats" -UseBasicParsing |
  Select-Object -ExpandProperty Content

If totalNodes and totalEdges stay unchanged across multiple content-rich sessions, run:

python scripts/repair_agentmemory_graph_windows.py --dry-run

Then apply with --all --restart-service only after checking the dry-run delta.

Common Non-Blocking Warnings

Do not confuse these with hook command failures:

• Shell snapshot not supported yet for PowerShell: Codex shell snapshot limitation, not an agentmemory hook failure.
• chatgpt authentication required for remote plugin catalog: remote plugin sync warning, not local hook execution.
• GitHub API rate limit exceeded: curated plugin sync warning, not agentmemory hook execution.
• resources/list or resources/templates/list unknown for agentmemory MCP: MCP capability warning, not hook capture failure.

Upgrade Warning

Upgrading or reinstalling @agentmemory/agentmemory can overwrite patched stop.mjs and dist\index.mjs. After an upgrade, rerun:

python scripts/repair_agentmemory_graph_windows.py --dry-run --patch-hooks --patch-service

If it reports pending patches, run the same command without --dry-run and restart agentmemory.