v3.0 was the foundation — MEMORIA architecture, structured facts, the BEAM jump from 35% to 65%. That was the hard part. v3.1 is the expansion layer. Not a rewrite. A growth spurt.
v3.1 started as a collection of fixes people actually hit in production. sqlite-vec int8 silently returning wrong results. Timezone drift in temporal recall. The Hermes plugin missing the bank parameter for multi-bank setups. And somewhere between fixing those, the community shipped features that turned into a proper minor release.
Shared Surface Memory
Until now every agent had its own private memory silo. If you wanted Agent A to know what Agent B learned, you either shared a database (dangerous) or re-ingested everything (expensive). Shared surface memory gives you a third option: a dedicated cross-agent surface with its own read/write/search/ delete/stats API. Each surface is fully isolated. Agents can share knowledge without sharing everything.
This unlocks multi-agent patterns that were awkward or impossible before. Orchestrator agents dump task summaries to a shared surface. Worker agents read from it. No database sharing. No cross-contamination.
Multilingual MEMORIA
MEMORIA's extraction pipelines were English-only at launch. Three contributors independently fixed that. WegoW shipped German detection and fixed language pattern collisions. Whishp added Russian regex extraction. And multilingual embedding auto-dim detection means Chinese, Japanese, French — any embedding model — now works without manual dimension configuration. The extraction layer auto-detects input language and routes to the right pattern set.
Custom Embedding Endpoints
flooryyyy shipped the ability to point Mnemosyne at any OpenAI-compatible embedding provider. Set `MNEMOSYNE_EMBEDDING_BASE_URL` and you're done. Jina model dimensions are auto-detected. Custom SSL certs supported. This means self-hosted embedding, alternative providers, or proxy setups — no more vendor lock-in for the vector pipeline.
The Fixes That Matter
Twenty-one fixes in this release. The ones that hurt most before:
- sqlite-vec int8 search. The int8 vector type requires `AND k=N` instead of `LIMIT`. Our vec_search was silently returning wrong results for anyone using int8 embeddings. Fixed.
- Timezone temporal recall. Queries with timezone-aware timestamps were off by an hour. Fixed.
- Working memory vectors. `remember()` was supposed to persist vectors on every call but the generation path was lazy. Fixed.
- Plugin bank parameter. All 6 Hermes plugin tool schemas now include the bank parameter. Multi-bank setups work.
- MULTILINGUAL_PATTERNS. Duplicate instruction keys and false positive German patterns across three separate rounds of fixes.
What's Next
v3.1 is the first release where community contributions outnumber internal ones by lines changed. That's the pattern I want to keep. The release after this one will focus on recall latency at 1M+ token scale and the BEAM benchmarking SDK — the tools that let you measure what your memory system actually does.
Until then.
$
pip install --upgrade mnemosyne-memory