Canonical copies of the aetnamem integration files for MemoryStackBench:

• targets/aetnamem.yaml — target manifest
• adapters/aetnamem.py — MemoryStackAdapter implementation

CI overlays these onto a fresh MemoryStackBench clone and requires 33/33 on seven_sins_v0_1 before a change can merge. Once these files are upstreamed into MemoryStackBench itself, the overlay step goes away and the benchmark repo stays the neutral referee.

graph_recall.py creates a temporary in-memory store, inserts synthetic records plus a two-hop relation, and compares bounded lexical and graph recall. It reports median/p95 latency, target rank, logged candidate count, and graph nodes visited:

python bench/graph_recall.py --records 10000 --iterations 25

This is a repeatable local probe, not a replacement for MemoryStackBench or a claim about production hardware. Use it to catch accidental O(N) retrieval or audit-payload growth while changing graph retrieval.

openclaw_memory/ contains a paired, fresh-session A/B benchmark of native MEMORY.md replay versus AetnaMem's bounded recall. It runs real OpenClaw turns through DeepSeek V4 Flash and preserves machine-readable token, cost, latency, accuracy, provenance, and retrieval evidence for technical reporting.