Summary

In a non-default agent session (agent:explore:main), the built-in memory_search tool appears to use the wrong/ambiguous current-session context and fails, even though the same agent's memory index and embeddings are healthy and openclaw memory search --agent explore works.

This makes memory recall unreliable for non-default agents and is especially confusing because different built-in session tools disagree about what the current session is.

Environment

• OpenClaw: 2026.6.6
• OS: macOS arm64
• Channel: webchat
• Affected agent: explore
• Memory provider for explore: custom provider M6 -> resolved as openai-compatible
• Embedding model: nvidia/nv-embed-v1

What I verified

1) The actual current conversation is in explore

Using session tools:

• sessions_history(sessionKey="current") returns:
  • sessionKey: "agent:explore:main"

So the live conversation really belongs to the explore agent.

2) session_status(current) disagrees

• session_status(sessionKey="current") reports:
  • Session: main

So at least one built-in tool resolves/displays current differently from sessions_history.

3) explore memory backend is healthy

openclaw memory status --deep --agent explore reports:

• Provider: openai-compatible (requested: M6)
• Model: nvidia/nv-embed-v1
• Indexed: 59/59 files · 673 chunks
• Embeddings: ready
• Semantic vectors: ready
• Vector dims: 4096

4) CLI memory search for explore works

Command:

openclaw memory search --agent explore --query "M6 embeddings" --max-results 5 --json

This successfully returns hits from memory/2026-06-15.md in the explore workspace.

5) Built-in memory_search tool fails in the same conversation

Calling the built-in tool with the same topic:

{"corpus":"memory","maxResults":5,"minScore":0.1,"query":"M6 embeddings"}

returns:

{
  "results": [],
  "disabled": true,
  "unavailable": true,
  "error": "memory_search timed out after 15s",
  "warning": "Memory search is unavailable due to an embedding/provider error.",
  "action": "Check embedding provider configuration and retry memory_search."
}

This error is misleading because the embedding/provider configuration for explore is already healthy and the CLI search succeeds.

Expected behavior

In a live session whose actual session key is agent:explore:main:

• memory_search should resolve the active agent as explore
• it should use the explore workspace/index/provider
• it should return the same class of results that openclaw memory search --agent explore returns
• session_status(current) should not collapse or misreport that session as plain main

Actual behavior

• sessions_history(current) says the current session is agent:explore:main
• session_status(current) says Session: main
• built-in memory_search fails/times out in that same session
• direct CLI search for --agent explore works fine

Why this looks like a bug

From the installed runtime code, the memory tool path appears intended to resolve the agent from the tool session key:

• resolveMemoryToolContext() in dist/tools-C_JRyDeX.js
• resolveSessionAgentIds() in dist/agent-scope-MrLta7Pq.js

resolveSessionAgentIds() should resolve agent:explore:main to agent explore.

So this looks like one of these:

1. memory_search is not receiving the correct agentSessionKey / runSessionKey in this live webchat route
2. the current-session alias/fallback machinery is collapsing a non-default agent main session into default main
3. session_status(current) and memory_search are not using the same current-session resolution rules

Reproduction outline

1. Create/use a non-default agent, e.g. explore
2. Configure working memory embeddings for that agent
3. Verify:
   • openclaw memory status --deep --agent explore
   • openclaw memory search --agent explore --query "..."
4. In a live conversation for that agent, check:
   • sessions_history(sessionKey="current")
   • session_status(sessionKey="current")
5. Call built-in memory_search
6. Observe:
   • sessions_history(current) resolves to agent:explore:main
   • session_status(current) may report main
   • built-in memory_search fails or times out while CLI --agent explore search works

Impact

This breaks reliable memory recall for non-default agents and makes diagnosis confusing, because the failure message suggests an embedding/provider problem even when the real issue is session/agent context resolution in the tool runtime.