openclaw/docs/plans/claude-mem-moltbot-fix.md

Claude-Mem Moltbot Plugin Fix Plan

Status Assessment

Current State:

• Plugin loads from: /Users/alexnewman/clawd/plugins/claude-mem/index.ts
• Observations: Working correctly (recorded via tool_result_persist hook)
• Prompts: Being recorded BUT with [message_id: ...] suffix appended
• MEMORY.md sync: Working (on session_start)
• Context injection: NOT implemented (only MEMORY.md file approach)

Root Cause: Moltbot's messaging system adds [message_id: UUID] to prompts for WhatsApp/Discord reactions and replies (see src/auto-reply/reply/body.ts:42). This metadata is being passed through to claude-mem without stripping.

Phase 1: Strip message_id from prompts

Problem

Line 304 in /Users/alexnewman/clawd/plugins/claude-mem/index.ts:

const result = await client.initSession(contentSessionId, event.prompt);

Sends the full prompt including [message_id: UUID].

Fix

Add a utility function to strip the message_id line before sending:

Copy pattern from: src/gateway/chat-sanitize.ts:18-38

const MESSAGE_ID_LINE = /^\s*\[message_id:\s*[^\]]+\]\s*$/im;

function stripMessageIdFromPrompt(text: string): string {
  if (!text.includes("[message_id:")) return text;
  const lines = text.split(/\r?\n/);
  const filtered = lines.filter((line) => !MESSAGE_ID_LINE.test(line));
  return filtered.join("\n").trim();
}

Implementation

Update before_agent_start hook (around line 297-310):

api.on("before_agent_start", async (event, ctx) => {
  if (!event.prompt || event.prompt.length < 10) return;
  if (!(await client.isHealthy())) return;

  const contentSessionId = getContentSessionId(ctx.sessionKey);

  // Strip [message_id: ...] metadata before recording
  const cleanPrompt = stripMessageIdFromPrompt(event.prompt);

  const result = await client.initSession(contentSessionId, cleanPrompt);
  if (result) {
    api.logger.debug?.(
      `claude-mem: session initialized (dbId: ${result.sessionDbId}, prompt#: ${result.promptNumber})`
    );
  }
});

Verification

1. Run moltbot with a test message
2. Check /api/prompts endpoint - prompts should NOT contain [message_id: ...]
3. Check claude-mem UI - prompts should display cleanly

---

Phase 2: (Optional) Add context injection via prependContext

The current plugin only syncs to MEMORY.md file. For real-time context injection (like the extensions/memory-claudemem approach), add:

api.on("before_agent_start", async (event, ctx) => {
  // ... existing prompt recording code ...

  // Optional: inject recent context
  const observations = await client.getObservations(5);
  if (observations.length === 0) return;

  const memoryContext = observations
    .map((obs) => `- [#${obs.id}] ${obs.title || obs.type}: ${obs.narrative || obs.text?.slice(0, 100)}`)
    .join("\n");

  return {
    prependContext: `<claude-mem-context>\nRecent memories:\n${memoryContext}\n</claude-mem-context>`,
  };
});

Note: This is optional since MEMORY.md file approach is already working.

---

Files to Modify

File	Change
/Users/alexnewman/clawd/plugins/claude-mem/index.ts	Add stripMessageIdFromPrompt() utility, update before_agent_start hook

---

Verification Checklist

• Prompts recorded without [message_id: ...] suffix
• Prompts display cleanly in claude-mem UI at localhost:37777
• Observations still being recorded correctly
• MEMORY.md still syncing on session_start
• No regressions in existing functionality