fix(memory): normalize non-canonical extraction payloads

[yeyitech]

Summary

• normalize several common memory-extraction payload shapes into the expected {"memories": [...]} form
• fall back to the stability parser when the primary JSON parser returns None
• add regression tests for bare objects, wrapped single objects, and nested wrapper keys

Why

Closes #1410.

Some smaller local / OpenAI-compatible models appear to return payloads that are structurally close to the expected schema but not exact, for example:

• a bare list of memory objects
• a bare single memory object
• {"memories": {...}} instead of {"memories": [{...}]}
• wrapper keys like items or nested data.memories

Before this change those responses could normalize to an empty result, causing memories_extracted to stay empty even when the model had clearly understood the conversation.

Validation

• python3 -m py_compile openviking/session/memory_extractor.py tests/session/test_memory_extractor_response_types.py
• python3 -m pytest -o addopts='' -p no:asyncio --noconftest tests/session/test_memory_extractor_response_types.py tests/session/test_memory_extractor_language.py -q

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis ✅ 1410 - Fully compliant Compliant requirements: Fix memory extraction returning empty dict when LLM executes successfully Handle non-canonical response shapes from smaller local models Add regression tests for various response formats

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🏅 Score: 92

🧪 PR contains tests

🔒 No security concerns identified

✅ No TODO sections

🔀 No multiple PR themes

⚡ Recommended focus areas for review Missing Observability Removed warnings for non-canonical response shapes (e.g., list instead of dict). While normalization is desired, losing these warnings may make it harder to diagnose model behavior changes. with telemetry.measure("memory.extract.stage.normalize_candidates"): data = parse_json_from_response(response) if data is None: stable_data, stable_error = parse_json_with_stability(response) if stable_data is not None: data = stable_data else: logger.warning( "Memory extraction stable parse failed: %s", stable_error, ) data = self._normalize_extraction_payload(data) Silent Filtering _coerce_sequence silently filters non-dict items from lists without logging. This could hide model output issues where some memories are lost. def _coerce_sequence(value) -> list: if isinstance(value, list): return [item for item in value if isinstance(item, dict)] if _is_memory_item(value): return [value] return []

[github-actions]

PR Code Suggestions ✨

No code suggestions found for the PR.

[yeyitech]

Updated based on the review suggestions. I restored lightweight observability for normalized non-canonical extraction payloads and added regression coverage for the new logging paths.