Release announcements and product updates, newest first. DBX v1.3.0 brings live vector migration, historical search, framework integrations, stronger isolation, and scoped support recovery. Opening this page clears the header badge. Follow the RSS feed or browse the changelog.
DBX v1.3.0 — memory you can evolve and operate New
Since v1.2.0, DBX has grown from its recall release into a fuller operating toolkit for per-tenant AI memory. You can change embeddings while current search stays available, inspect retained retrieval history, connect application frameworks, and diagnose a tenant through a scoped support interface.
Evolve memory.VMIGRATE START/ADD/SWAP/CANCEL builds and promotes a shadow index with WAL-durable lifecycle operations. VSEARCH ... AS_OF reconstructs retained vector history for retrieval audits and incident reproduction.
Connect your application. LangChain and LlamaIndex adapters, bounded batch ingest, RAG examples, refreshed dashboard views, and command-string or argument-array HTTP queries.
Strengthen the boundary. Linux worker seals cover all worker threads, seccomp restricts network socket creation, and WAL replicas use mutual authentication and AES-256-GCM encrypted frames.
Support one tenant at a time. A separately installed extension reads allowlisted summaries. An optional, distinct wake capability can wake a hibernated tenant with a durable pre-action audit, cooldown, and independent post-check.
Exercise the decisions. The support curriculum includes 31 casebook entries, 43 fault/contract tests, and 1,340 repeatable synthetic diagnostic and recovery-policy drills.
Build on examples. Legal Auditor and Jurix demonstrate application workflows around tenant-scoped memory and retrieval.
Migration requires application-side re-embedding and temporary storage. Historical search is bounded by retained WAL. Replication remains asynchronous, and migration lifecycle records are not yet a replicated protocol. Support is opt-in and cannot repair arbitrary defects or restore, delete, or reconfigure customer data.
The single-node profile remains 100 tenants per node, 100k vectors per tenant, and durable strings plus vectors. Existing performance figures retain their original certification date and workload.
DBX now connects two operational needs for per-tenant AI memory: changing embeddings without taking current search offline, and reproducing what a tenant’s vector index looked like at an earlier point in time.
VMIGRATE builds a complete shadow index while the current index remains searchable. Stream caller-generated vectors with VMIGRATE ADD, promote a non-empty target with VMIGRATE SWAP, or cancel without changing the live index.
VSEARCH ... AS_OF <unix-nanoseconds> replays retained vector WAL history into a temporary index for audit trails, RAG incident reproduction, and compliance investigations.
These workflows are tested for overwrite, delete, cancellation, concurrent workloads, and load. Honest limits: migration requires caller-side re-embedding and temporary storage; its lifecycle is not yet a replicated WAL protocol. Historical search is slower than current search and is bounded by WAL retention.
v1.1.0 was Isolation Kernel: one worker per tenant, Unix sockets, HELLO / INFO, and a VSEARCH-shaped recall path. v1.2.0 keeps that kernel and finishes the recall surface.
VSEARCH — caller vector, cosine over that tenant’s HNSW. Trailing MIN_SCORE, EF, SPACE.
VSIM — neighbors of a stored id, self excluded. “More like this memory” without re-embedding.
VFUSE — late fusion across named SPACE graphs with caller WEIGHTS. Missing ids score 0. Weighted cosine sum, not a multimodal model.
Python helpers: vsearch / vsim / vfuse, plus TenantMemory.similar and TenantMemory.fuse. Playground tabs: Semantic / Similar-to-id / Multimodal. The public site adds a toy cosine bench, density slider, fit quiz, DEK-shred sketch, clickable video chapters, Updates badge, narrated two-minute clip, and a labeled shared-cluster vs DBX isolation sketch.
Is it better than v1.1.0? Yes as a product an agent uses — the verbs, the playground, and the honest site are what were missing. No as a kernel: certified ANN p50 is still 2.304 ms, Isolation Kernel strict idle RSS is still ~14–17 MiB per worker, still 100 tenants per node and 100k vectors per tenant. Not a Qdrant competitor. GitHub Release v1.2.0 · GHCR :v1.2.0.
Engine detail for the v1.2.0 cut. Embeddings stay with the caller. DBX does not run CLIP or MiniLM. A named space is another mmap/HNSW file in the same tenant directory — backup, quota, and shred still hit one customer.
Linux strict seals each tenant as its own process: Landlock-restricted filesystem, encrypted WAL / checkpoints / vector ids / HNSW graph, Unix sockets only the orchestrator can open. Idle RSS in the 14–17 MiB band.
Honest limits: SQ8 .vec rows stay mmap plaintext so idle tenants live in page cache. Put the data dir on LUKS or fscrypt. make run-dev is still inprocess — do not ship that as Isolation Kernel. We are not SOC 2 certified.
A recorded demo of the public site and the dashboard compiled into dbx-orchestrator: login, provision, mint a writer, console SET/GET/VADD, explorer, backup. Engineer voiceover over a 140 BPM tech bed. The dashboard itself is not on GitHub Pages.