feat(memory): post-final-answer memory extraction via standalone LLM call

[JasonW404]

Summary

Refactors the memory service execution logic to decouple memory extraction from the react loop.

Problem

Previously, memories were only stored when the LLM called StoreMemoryTool inline during the agent loop — a model-directed approach that was unreliable. The LLM often forgot or skipped the call at the final answer step.

Solution

Two complementary mechanisms:

1. Restrict StoreMemoryTool to action steps only

• Updated system prompt (zh/en) to prohibit store_memory during the final answer step
• Updated ToolConfig description to match the new policy
• StoreMemoryTool now only fires during intermediate action steps for process-level observations

2. Post-final-answer extraction pipeline

• New FaMemoryExtractor module (backend/services/fa_memory_extractor.py)
• New extraction prompt YAML (backend/prompts/fa_memory_extraction_en.yaml)
• Wired into agent_service.py finally block as fire-and-forget asyncio.create_task
• Non-blocking — never delays the user response
• Uses tenant LLM to parse <memory-item> tags from the final answer
• Stores extracted items via MemoryService with layer=AGENT, memory_type=SHORT_TERM

3. Bug fix

• memory_type enum normalization in _backend_store_hook used isinstance(x, MemoryLayer) instead of hasattr(x, 'value')

Files Changed

File	Change
backend/services/fa_memory_extractor.py	New — FaMemoryExtractor class + ExtractionResult dataclass
backend/prompts/fa_memory_extraction_en.yaml	New — extraction prompt with <memory-item> output format
test/backend/services/test_fa_memory_extractor.py	New — 29 unit tests
backend/utils/memory_tool_prompt.py	Restrict store_memory to action steps
backend/agents/create_agent_info.py	Update ToolConfig description
backend/services/memory_backend_adapter.py	Add build_memory_service_for_fa_extraction() + fix enum bug
backend/services/agent_service.py	Capture final answer + schedule extraction task
test/backend/services/test_memory_backend_adapter.py	+9 tests (enum normalization, embedding skip, defaults)

Test Results

• 29/29 new FaMemoryExtractor tests pass
• 16/16 memory_backend_adapter tests pass (7 existing + 9 new)
• 71/71 SDK memory tests pass (no regressions)
• End-to-end verified through real chat API pipeline: 3 memories successfully extracted and stored from a single conversation

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.