openclaw - 💡(How to fix) Fix Memory plugin auto-recall embeds full conversation context, causing context length errors and timeouts [1 participants]

Summary

Auto-recall for memory plugins (specifically openclaw-mem0 in open-source mode) fails on conversations with large context because the full conversation is passed to the embedding model instead of a focused query.

Environment

• OpenClaw: 2026.4.15
• Plugin: openclaw-mem0 1.0.10
• Mode: open-source (self-hosted)
• Embedding model: Ollama nomic-embed-text (~8K token context limit)
• Vector store: Qdrant (localhost)
• OS: Windows 11

Observed Behavior

1. Context length error when conversation includes large content (PDFs, long code blocks):

   [plugins] openclaw-mem0: recall failed: ResponseError: the input length exceeds the context length

2. 8-second timeout on moderate-to-large conversations:

   [plugins] openclaw-mem0: recall timed out after 8000ms, skipping

3. Normal short conversations work fine — recall succeeds and memories are injected.

Expected Behavior

• Auto-recall should extract a focused query from the conversation (e.g., last user message, ~500 tokens max) rather than embedding the entire context
• Large attachments, PDFs, or code blocks in the conversation should not affect recall
• The embedding input should be deterministic and bounded

Root Causes (Analysis)

1. Full context passed to embedding: The plugin receives the entire conversation context for recall, then attempts to embed it. This exceeds embedding model limits.

2. Hardcoded 8s timeout: The 8000ms timeout is not configurable. Even when embedding succeeds, large contexts can timeout.

3. No graceful degradation: When embedding fails, recall fails entirely rather than falling back to a truncated query.

Suggested Fixes

1. Bound the recall query: OpenClaw should pass a bounded context slice to memory plugins (e.g., last N tokens, configurable)

2. Expose recallTimeoutMs config: Allow users to increase timeout for slower embedding providers

3. Plugin contract clarification: Document whether plugins are responsible for truncation, or whether OpenClaw guarantees bounded input

Workaround

None currently. Auto-recall fails silently on large contexts; users must use manual memory_search tool calls.

Reproduction Steps

1. Configure openclaw-mem0 in open-source mode with Ollama embeddings
2. Enable autoRecall: true
3. Send a message with a large PDF attachment or paste a long document
4. Observe gateway logs showing recall failure

Related

This may also need changes in the openclaw-mem0 plugin to truncate before embedding, but the architectural decision of what gets passed to plugins is OpenClaw's responsibility.

Happy to provide additional logs or test configurations if helpful.