Run full-scale LongMemEval and LoCoMo benchmarks

[lightcone0]

Context

db0's current benchmark scores are based on limited samples:

• LongMemEval: 80.0% on 50 questions (of 500 total)
• LoCoMo: 76.9% on 199 queries from 1 sample (of 10 samples, ~1986 queries total)

These are promising but not directly comparable to published scores from other systems that run the full datasets.

What's needed

1. Full LongMemEval run — all 500 questions, multiple profiles (conversational, high-recall, knowledge-base), with Gemini and OpenAI embeddings
2. Full LoCoMo run — all 10 samples (~1986 queries), multiple ingestion modes
3. Category breakdown — especially temporal reasoning, knowledge-update, and multi-session categories where architecture differences matter most
4. Profile comparison — demonstrate that the right profile matters: run the same benchmark with conversational vs. high-recall vs. knowledge-base profiles

Expected outcome

• Validated scores comparable to published results from Hindsight (91.4%), Zep (71.2%), Mem0 (49-68.5%)
• Data showing profile impact on different query categories
• Identification of specific weak spots (likely temporal reasoning) to guide Add BM25/keyword search as a retrieval strategy #10, Add bi-temporal model for temporal reasoning queries #11, Auto-extract entity graph from conversations #12

Practical notes

• Full LongMemEval with Gemini embeddings + judge takes ~4-8 hours
• Full LoCoMo with 10 samples takes ~2-3 hours
• Both require GEMINI_API_KEY or OPENAI_API_KEY
• Results should be updated in packages/benchmark/README.md