This release fixes issues for SREs watching stability and regressions.
Published 3mo
MCP Data & Storage
✓ No known CVEs patched
✓ No known CVEs patched in this version
Summary
AI summaryFixed attention‑mask corruption in batch mean pooling that caused embedding errors on padded tokens.
Full changelog
🐛 Bug Fixes (13/16 from audit)
Wave 1 — Zero-Risk (7 fixes)
- BUG-2, BUG-15: NaN-safe sort in
search.rsandrrf.rs - BUG-4: Dockerfile HEALTHCHECK
start-period60s → 120s - BUG-8: blake3-based ID generation replacing weak
nanos XOR pid - BUG-9: Queue metrics drift prevention —
inc_queue()aftersend() - BUG-10:
invalidate()storessuperseded_byparameter - BUG-13: Embedding cache purges stale entries on model change
Wave 2 — Medium-Scope (5 fixes)
- BUG-3: Runtime
.expect()→ graceful error withSTATUS_ERROR - BUG-6:
embed_for_record()passes actual chunk ID - BUG-7: File-level hash comparison via
blake3::Hasherfor incremental indexing - BUG-12: Watcher cancel channel prevents post-stop debounce callbacks
- BUG-16:
memory_typeparse validation with error instead of silent drop
Wave 3 — Critical (1 fix)
- BUG-1: Attention mask for batch mean pooling — padding tokens no longer corrupt embeddings
✅ Verification
cargo build --profile fast— all 3 wave gates passedcargo test— 69 passed, 0 failed
Not Fixed (by design)
- BUG-5: dead config field (cosmetic)
- BUG-11: noop flush (architectural)
- BUG-14: schema dimension matches current model
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About pomazanbohdan/memory-mcp-1file
A self-contained Memory server with single-binary architecture (embedded DB & models, no dependencies). Provides persistent semantic and graph-based memory for AI agents.
Related context
Beta — feedback welcome: [email protected]