Bug

The openclaw-mem0 plugin maps run_id to the OpenClaw sessionKey (e.g. agent:main:telegram:direct:5263871946), which is a durable channel identifier that survives /new and /reset commands. This means:

1. Session memories are never isolated per conversation — all auto-captured memories share the same run_id across session resets
2. memory_list(scope=session) returns all auto-captured memories ever, not just the current conversation
3. Session memory pool grows unbounded (168+ in my case after a few days)

Expected behavior

/new should generate a fresh run_id so that session-scoped memories are isolated to the current conversation, matching Mem0's intended semantics:

Additional bug found

Auto-capture uses ctx.agentId (e.g. "main") as user_id, but the memory tools (memory_list, memory_search, memory_store) fall back to cfg.userId (e.g. "default"). This causes tools to miss all auto-captured memories entirely. I patched this locally by adding currentAgentUserId fallback to the tools.

Proposed fix

1. Generate a new run_id (UUID or incrementing counter) on /new//reset, pass it to the plugin instead of sessionKey
2. Consistently use user_id for the human identity across auto-capture and tools
3. Use agent_id for the agent name (currently unused)
4. (Plugin-side) Fix currentAgentUserId fallback in tool handlers (I have a working patch)

Environment

• OpenClaw 2026.2.26
• openclaw-mem0 plugin v0.1.2 (OSS mode)
• Qdrant vector store
• Mem0 SDK 2.2.1

/cc @gevertex