You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Create a standalone benchmark suite (basic-memory-bench) for evaluating retrieval quality across BM deployments using academic datasets. Designed to be publicly shareable, runnable by anyone, and integrated into CI.
Why
Internal quality tracking — benchmark before/after every release
Cloud vs Local comparison — validate that Cloud's better embeddings produce better retrieval
Public credibility — reproducible numbers on academic benchmarks
Marketing — "we benchmark in the open" content
Competitive positioning — compare against Mem0/Supermemory on the same datasets
Current Results (from prototype in openclaw-basic-memory plugin)
Full LoCoMo benchmark — 1,982 queries across 10 conversations:
Metric
BM Local (v0.18.5)
Recall@5
76.4%
Recall@10
85.5%
MRR
0.658
Content Hit Rate
25.4%
Mean Latency
1,063ms
By category:
Category
N
R@5
open_domain
841
86.6%
multi_hop
321
84.1%
adversarial
446
67.0%
temporal
92
59.1%
single_hop
282
57.7%
Architecture
Python (not TS) — same ecosystem as BM, uses BM's importer framework
Summary
Create a standalone benchmark suite (
basic-memory-bench) for evaluating retrieval quality across BM deployments using academic datasets. Designed to be publicly shareable, runnable by anyone, and integrated into CI.Why
Current Results (from prototype in openclaw-basic-memory plugin)
Full LoCoMo benchmark — 1,982 queries across 10 conversations:
By category:
Architecture
EntityMarkdownjust full-locomoruns everything, fail if recall drops >2%Datasets
Known Improvement Opportunities
From analysis of the 375 failures:
Implementation Phases
Reference
drafts/spec-benchmark-suite.mdin openclaw workspacebenchmark/in openclaw-basic-memory repo