Summary

Add defensive type checking in memory extraction to handle LLM responses that return a list instead of a dict. This fixes crashes when using local Ollama models.

Why this matters

When using OpenViking with local Ollama models (e.g., llama3.1:latest), memory extraction fails because parse_json_from_response() can return a list [{...}] instead of the expected {"memories": [{...}]}. The subsequent data.get("memories", []) call throws AttributeError since lists don't have .get().

• #605 - Full repro with Ollama config (bge-m3 + llama3.1)
• The existing codebase already uses isinstance(data, dict) checks in the merge bundle path (line 559), but the extraction path at line 296-301 lacks this guard

Changes

In openviking/session/memory_extractor.py, after parse_json_from_response():

• If data is a list, wrap it as {"memories": data} (treat the list as the memories array directly)
• If data is neither dict nor list, log a warning and fall back to empty dict
• Existing dict-format responses pass through unchanged

Testing

Added tests/session/test_memory_extractor_response_types.py with 6 test cases:

• Dict response passthrough (existing behavior preserved)
• List response wrapping
• String/None/int fallback to empty dict
• Empty list handling

Fixes #605

This contribution was developed with AI assistance (Claude Code).