chore: upgrade mem0ai to 2.0.5 and adapt provider to the 2.0 search API

[groksrc]

Summary

The mem0-local provider was pinned to mem0ai==1.0.0 (November-era). mem0 has since shipped 2.0 (currently 2.0.5), whose Memory.search API is a breaking change:

• entity scoping moved from a top-level user_id= kwarg to filters={"user_id": ...}
• limit= became top_k=

The 2.0 API rejects the old call shape outright (_reject_top_level_entity_params), so the provider could not run against current mem0 at all — meaning headline comparisons were silently testing a major version behind mem0's current state.

Changes

• Bump pin to mem0ai[nlp]==2.0.5 (the nlp extra activates mem0's spaCy lemma/keyword pipeline instead of running degraded with a warning)
• Update the single Memory.search call in mem0_local.py to the 2.0 signature; add/delete_all are unchanged in 2.0

Fairness / reproducibility impact

• mem0 runs at its current released version with out-of-the-box defaults (score threshold 0.1, rerank off) — no provider-specific tuning, per the fairness contract.
• Operators need spaCy's en_core_web_sm model present; mem0 auto-downloads it where pip is available. In uv-managed venvs (no pip), preinstall with uv pip install <en_core_web_sm wheel URL>.
• Gotcha worth knowing for run hygiene: mem0 2.0 silently disables its v3 hybrid (BM25) search if the local Qdrant collection at /tmp/qdrant was created by an earlier version — wipe /tmp/qdrant and ~/.mem0 before fresh comparison runs.

Validation

• uv run ruff check . and uv run pytest tests pass
• Full LoCoMo run (1,986 queries) completed with mem0-local state ok on 2.0.5

🤖 Generated with Claude Code