Skip to content
Memoturn

Roadmap

Memoturn is at the architecture + working prototype stage. All planned prototype milestones are built and tested: 67 cross-crate integration tests covering the memory layer, HTTP-level hardening regressions, and true multi-node distribution against a real etcd (gated on ETCD_ENDPOINTS). The stack is proven on Kubernetes — the Helm chart deploys to kind with MinIO and auth, all HTTP benchmarks pass, and the chaos test holds: deleting the data-plane pod yields a fresh pod with no PersistentVolume serving the same data with the same token in about 15 seconds. See deployment.

Performance numbers on this page and in the repository are p50 prototype measurements (single node, in-process object store), reproducible with cargo run --release -p memoturn-bench.

Shipped surface

Section titled “Shipped surface”
  • Typed agent memory — namespace > profile > memory, where a profile is one database. Idempotent content-addressed ingest, supersession by topic key, hybrid recall (keyword + topic
    • vector, reciprocal-rank fused), sessions, forget. See memories and recall.
  • Opt-in node services — server-side extraction (/extract, 503 when unconfigured) and auto-embedding, both outside the write path.
  • Multi-model substrate — documents on JSONB, KV with TTL, vectors, the transcript layer, and the SQL escape hatch, all in one database per profile. See data model.
  • Branching — O(1) copy-on-write forks, checkpoints, rewind, and auto-expiring burner branches. See branching.
  • Distribution — object storage as the source of truth, single writer per database with lease + epoch fencing, write forwarding, replica reads with txid/min_txid — integration- tested multi-node against a real etcd. See architecture and consistency.
  • Durability & GC — opt-in Durable write mode (MEMOTURN_DURABILITY=durable, per-request Memoturn-Durability header) that acks only after segment ship + manifest CAS, and a refcounted object GC that is copy-on-write safe across branch manifests. See consistency.
  • Access surfaces — REST API, CLI, TypeScript SDK, Python SDK, and MCP server.
  • Operations & hardening — fail-closed auth (per-database and namespace JWTs, platform key, token revocation on database deletion), the SQL guard, request-surface limits, env-var node configuration, and a hardened Helm chart (non-root read-only pods, NetworkPolicy, no K8s API token, etcd required for multi-replica). See security and deployment.

Deferred work

Section titled “Deferred work”

These items are tracked in the architecture decision records at github.com/memoturn/memoturn (docs/adr/). No dates are attached; the ordering below reflects dependency, not commitment.

Storage and engine

Section titled “Storage and engine”
  • Lazy page-fault VFS for databases over 16 MB. The prototype restores whole databases on cold wake, which is the right trade for tiny per-profile databases; larger databases need page-granular faulting from object storage.
  • Point-in-time-recovery retention policy. Branch, rewind, and refcounted GC of unreferenced objects exist; a retention/compaction policy for the immutable segment log does not yet.

Branching and lifecycle

Section titled “Branching and lifecycle”
  • Three-way branch merge. Out of scope for v1 by decision: agents fork-and-promote or fork-and-discard. Merge matters once branches host long-lived divergent work.
  • Whole-database TTL for ephemeral profiles on multi-node clusters. Burner branches already expire; expiring an entire profile database needs cluster-wide sweep coordination.

Distribution and operations

Section titled “Distribution and operations”
  • Cross-region replication — bucket cross-region replication plus remote read replicas.
  • Per-tenant KMS encryption of segments and manifests.
  • Kubernetes operator and CRDs, replacing chart-level wiring for fleet operations. See scaling and observability.
  • gRPC transport for the node mesh. HTTP suffices at prototype scale.
  • Billing pipeline beyond the current usage counters.
  • Namespace as a first-class catalog column. Profiles are databases named {ns}--{profile}; the convention upgrades to a catalog column with a one-time split.

Assistant

Section titled “Assistant”
  • The full built-in assistant — natural-language query, schema and index advice, and an ops copilot. The design is shipped (docs/architecture/06-mcp-and-assistant.md); the CLI’s ask command is a stub today and says so. Recall answer synthesis (turning recalled memories into a prose answer) lands here too, in the control plane rather than the data path.

Sequencing

Section titled “Sequencing”

The data plane is deliberately small and the invariants are fixed: object storage as the source of truth, single writer with epoch fencing, txid on every read. Deferred items either extend the manifest layer (VFS, GC, merge, TTL), the control plane (operator, billing, catalog, assistant), or the deployment story (cross-region, KMS) — none change the wire protocol in api-rest. Items land when they are provable with the same end-to-end rigor as the existing 67-test integration suite, not before.