WaveMind v2.2.2 - production load benchmarks

WaveMind v2.2.2

This release packages the current production-scale benchmark work and the launch material for WaveMind as a dynamic memory layer.

Highlights

• Added checked-in production load benchmark artifacts for local and service-backed retrieval.
• Documented the current 100k and 1M Qdrant service profile.
• Refreshed the launch kit with a clearer public story: WaveMind is not just vector search, it is memory with TTL, hotness, decay, corrections, namespaces, and an auditable SQLite source of truth.
• Fixed and refreshed the Russian launch post material.
• Verified the package with the full local test suite, build, and package validation before tagging.

Current benchmark snapshot

Profile	Recall@10	Avg query latency
Qdrant service, 100k memories	1.000	10.76 ms
Qdrant service, 1M memories	0.506	45.81 ms

The 1M run is intentionally included as an honest tuning target, not as a finished production claim. The next technical focus is improving large-scale recall while keeping latency within a production-friendly range.

Why it matters

Static vector databases are good at similarity search. WaveMind is designed for long-lived application and agent memory: newer facts can override stale ones, old memories can decay, namespaces keep tenants separate, and recall signals can change future retrieval.

Links

• Benchmark report: docs/BENCHMARK_REPORT.md
• Launch kit: docs/LAUNCH_KIT.md
• Benchmark brief: docs/BENCHMARK_BRIEF.md
• Full changelog: v2.2.1...v2.2.2

Thanks to @aouxwoux for the first external contribution in this release cycle.