Memoturn gives every user, team, or agent persona an isolated memory profile all of their agents share — typed memories with supersession, hybrid recall, and a tiny per-profile database you can checkpoint, fork, and rewind. Self-hosted, on any Kubernetes.
why memoturn
Agents remember in app code today: a vector store here, a KV cache there, transcripts somewhere else — and nothing forgets correctly.
Agent memory is not a vector store with extra steps. Memoturn moves memory into the database: typed records that supersede instead of duplicate, recall that fuses keyword, topic, and vector signals, and one profile boundary no data ever crosses — in one unit that replicates, branches, and rewinds together.
what you get
Typed memories, hybrid recall, and branching aren't an app-side library — they're database semantics you can inspect, self-host, and operate. Open source, top to bottom.
shared
Each user, team, or persona gets an isolated memory profile that all of their agents read and write. Isolation is structural — namespace > profile > memory — no data ever crosses a profile boundary, and deleting a profile revokes its stale write tokens.
evolving
Typed memories — facts, events, instructions, tasks — with supersession by topic: a new fact replaces the old one on the same topic and keeps the history. Ingest is idempotent; forget is explicit.
relevant
Hybrid recall fuses full-text, topic, and vector search with reciprocal-rank fusion and channel attribution — and returns empty when nothing matches, instead of guessing.
also in the database
database per profile
Every profile is its own tiny libSQL database — provisioned in microseconds, hibernating to object storage at near-zero cost when idle, holding memories, documents, KV, vectors, and transcripts in one unit.
burner branches
O(1) copy-on-write forks of an agent's whole memory. Checkpoint before a risky run, rewind if it learned garbage, branch a session and let it expire — branching is a manifest write, not a data copy.
multi-model API
Document collections on JSONB, KV namespaces with TTL, vector search, and a SQL escape hatch — one consistency model across all of it, with every read disclosing its `txid`.
object storage is truth
Committed transactions ship as immutable segments to object storage — opt into durable mode and writes ack only after the segment lands. Nodes are disposable, replicas stream the log, and writer leases with epoch fencing make zombie writers harmless.
extraction & embeddings, opt-in
Bring your own pipeline by default, or opt in per node: server-side extraction distills raw transcripts into typed memories, and auto-embedding adds the vector channel to bare-text ingest — both outside the write path.
self-hosted
A Rust data plane and one Helm chart hardened by default — non-root, read-only rootfs, locked-down egress — on any Kubernetes (EKS, GKE, AKS, or your own metal) with any S3-compatible object store. MCP server, CLI, and TypeScript/Python SDKs included.
what you build
The same profile, supersession, recall, and branching primitives back a range of agents — from personal assistants to multi-tenant fleets — all on infrastructure you own.
01
Assistants that actually remember
Facts and preferences persist across every session and every agent the user touches — and a new fact about their diet supersedes the old one instead of contradicting it.
02
Multi-agent fleets with shared state
Planner, researcher, and executor agents read and write one profile — each contribution attributed, nothing leaking between users, because the boundary is a database file.
03
Safe experimentation on live memory
Snapshot an agent's mind before a risky run, let it learn, and rewind or promote — burner branches are O(1), copy-on-write, and auto-expire.
04
Per-session scratch state
Task memories attach to a session and expire on TTL; the verbatim transcript layer keeps every turn searchable until you end the session — no cleanup job.
05
Tool-native memory over MCP
Agents ingest and recall memory as MCP tools — no SDK in the loop required — with the same namespace tokens and profile boundaries as every other surface.
06
SaaS agents at tenant scale
Millions of tiny databases on disposable nodes: one per user, near-zero cost idle, single-writer SQL semantics with failover under fifteen seconds.
ready when you are
Open source, Apache-2.0. Run `memoturnd` for a local node, or one Helm chart on any Kubernetes for production.