Most RAG setups are quietly a small distributed system: a vector database, a full-text index, an embedding service, and a metadata store, each with its own deploy step and failure mode. The bet here is that an agent's entire long-term memory can instead live in one immutable file you copy and version like a SQLite database — content, embeddings, search structures, and metadata bundled into append-only "Smart Frames" with no sidecar .wal, .lock, or .shm files and no server to run.
What Sets It Apart
- Single-file and serverless: a
.mv2artifact is something you can ship, diff, and version-control, so memory stops being infrastructure you operate and becomes a file you move around. - Hybrid retrieval in one container: BM25 full-text, HNSW vector similarity, and temporal/multi-hop reasoning sit in the same file, alongside ingestion for PDFs, CLIP image embeddings, and Whisper audio — one format instead of a glued-together stack.
- Append-only Smart Frames are immutable and checksummed, which buys two things at once: timeline queries that show how knowledge evolved, and crash safety because committed frames can't be half-written.
- Designed for local-first speed and offline use — it reports sub-millisecond local reads and benchmark gains on LoCoMo long-horizon recall, and stays model-agnostic.
Great Fit / Look Elsewhere
Great fit if you want agent memory you can ship as a file: local-first or offline apps, reproducible evaluation, or replacing a multi-service RAG stack you'd rather not run. Look elsewhere if you need concurrent multi-writer access at database scale, or your data already lives in a managed vector store you're happy with — a single-file format trades horizontal scale and write concurrency for portability. The headline latency and accuracy figures are self-reported, so validate them on your own corpus before committing.
