open source

Memoturn

Every user, team, or agent persona gets an isolated memory profile that all of their agents share — typed memories with supersession, hybrid recall, and a per-profile database you can checkpoint, fork, and rewind. Self-hosted, on any Kubernetes.

Read the quickstart See use cases View on GitHub

run it

One ingest call, durable memory

Memoturn is an open-source, distributed agent-memory database. Every agent serving a user is a client of the same profile: facts supersede by topic with history kept, ingest is idempotent, and recall fuses keyword, topic, and vector channels — empty when nothing matches, never padded.

memoturn memory ingest acme alice --type fact --topic user.diet \
--summary "vegetarian since 2024"
memoturn memory recall acme alice "what can this user eat?"

A Rust data plane embeds libSQL and gives each profile its own tiny database — provisioned in microseconds, hibernating to object storage at near-zero cost when idle, holding memories, documents, KV, vectors, and transcripts in one unit that replicates, branches, and rewinds together. Read the architecture overview for the full design.

connect Tool-native over MCP

Agents ingest and recall memory as MCP tools — memory_ingest, memory_recall, memory_forget.

experiment Burner branches

Checkpoint, fork, and rewind a whole profile in O(1) — ephemeral forks auto-expire.

operate Self-hosted

One Helm chart on any Kubernetes; object storage is the source of truth, nodes are disposable.

open source · Apache-2.0

Give your agents memory on infrastructure you own.

Typed memories with supersession, hybrid recall, and O(1) burner branches — database semantics, not app-side glue. Run a local node with memoturnd, or deploy the Helm chart to any Kubernetes with any S3-compatible object store. See Deployment for the path.

Get the source Quickstart