Add benchmark suite for memory quality evaluation

[rohitg00]

Problem

No published benchmarks or evaluation scores. Cannot demonstrate quality vs alternatives. In 2026, published benchmark results are the primary credibility signal for memory systems.

Proposed Solution

1. Implement evaluation scripts against standard benchmarks (LOCOMO, LongMemEval, AMA-Bench)
2. Create an internal eval harness that measures:
   • Retrieval precision/recall at k
   • Compression quality scores (already have `scoreCompression`)
   • End-to-end accuracy (store → retrieve → use in context)
   • Latency percentiles (P50, P95, P99)
3. Add benchmark results to README
4. Run benchmarks in CI to catch regressions

Impact

Establishes credibility and enables data-driven optimization decisions.

[rohitg00]

Resolved in v0.6.0 (PR #76). Full benchmark suite with labeled ground truth: benchmark/quality-eval.ts (20 queries, Recall@K/Precision@K/NDCG/MRR), benchmark/scale-eval.ts (240-50K observations), benchmark/real-embeddings-eval.ts (Xenova local embeddings). Reports: QUALITY.md, SCALE.md, REAL-EMBEDDINGS.md.