rUv b395d12559 feat: ADR-088 LongMemEval benchmark harness for AgentDB (#1566) ...

Full benchmark implementation to evaluate AgentDB memory system against
the LongMemEval standard (ICLR 2025, 500 questions, 6 categories).

Includes:
- ADR-088 with competitive landscape analysis (MemPalace 96.6%,
  OMEGA 95.4%, Observational Memory 94.87%)
- Benchmark harness with 4 modes: raw, hybrid, full, baseline
- AgentDB adapter (HNSW + ONNX) and baseline adapter (cosine)
- Dataset download scripts (HuggingFace)
- Per-category scoring, latency percentiles, storage metrics
- Honesty protocol: no test-set tuning, full category breakdown

Motivated by MemPalace's claimed "perfect score" — we need a real
apples-to-apples comparison to establish AgentDB's position.

Co-authored-by: Reuven <cohen@ruv-mac-mini.local>

2026-04-08 09:59:57 -05:00

..

download-dataset.sh

feat: ADR-088 LongMemEval benchmark harness for AgentDB (#1566)

2026-04-08 09:59:57 -05:00

run-benchmark.sh

feat: ADR-088 LongMemEval benchmark harness for AgentDB (#1566)

2026-04-08 09:59:57 -05:00