Long-term memory ingests full history every turn and stores chit-chat

[SamariddinS]

Problem

MemoryService.add() is invoked at the end of every chat turn with the
entire LangGraph message history (response["messages"] /
state.values["messages"]). Two failure modes stack:

1. Repeated re-extraction. Mem0's fact extractor runs against every
   prior turn on every call, so the same messages are paid for N times
   and near-duplicate rows slip past dedup into pgvector.
2. Permissive default prompt. Mem0's out-of-the-box extractor
   captures greetings, acknowledgements, transient task instructions,
   and assistant chit-chat — exactly the noise we don't want persisted.

Result: longterm_memory fills with rows like "2 + 2",
"Assistant offered to help", duplicate restatements of the same
preference, etc. Memory search quality degrades, token spend on the
extractor grows linearly with session length.

Reproduction

1. Start a session, send "what is the sum of 2 + 2" → a memory row is written "Asked how much is 2 + 2".
2. Send any follow-up → the extractor re-processes the sum plus
   the new turn; more near-duplicate rows land in pgvector.
3. Inspect longterm_memory rows for the user — pointless entries
   dominate the legitimate facts.

Impact

• Polluted long-term memory → worse retrieval relevance on
  memory_service.search.
• Unnecessary LLM spend on LONG_TERM_MEMORY_MODEL (extractor re-runs
  the full history).
• Cost scales O(turns²) over a session instead of O(turns).

Proposed fix in PR #66