[eric7h]

Summary

VikingDBLTMBackend.search_memory took data.result_list and returned a single JSON string whose parts[0].text was str(result_list) — the Python list repr of every hit. LongTermMemory then turned that into one unusable MemoryEntry, discarding per-hit memory_info (summary / profile / event fields) and the top_k shape.

Fix

• Emit one {"role":"user","parts":[{"text": ...}]} entry per result_list item.
• Prefer readable fields from memory_info (summary, user_profile, original_messages, question, answer, content, text); otherwise serialize memory_info.

How to reproduce

# Fake collection returning two hits with memory_info.summary:
backend.search_memory(user_id="user-42", query="tea", top_k=5)
# Before (main): 1 string, text starts with "[{'id': 'm1', ...}]"
# After: 2 strings with texts "likes tea" and "lives in Shanghai"

New tests in tests/test_vikingdb_memory_backend.py fail on main and pass with this change. Full file: 22 passed.

Test plan

• pytest tests/test_vikingdb_memory_backend.py — 22 passed
• pre-commit (ruff check/format, gitleaks) on touched files