memory for code agents, with provenance

Stop re-teaching agents the same repo decisions.

membl records durable engineering facts from agent sessions, keeps evidence and status with each fact, and recalls only the active project-scoped context the next agent needs. The core engine runs locally; Cloud adds the managed dashboard and team review flow.

Session historyLong, chronological, and expensive to re-send. Good for audit, bad for targeted recall.
Vector memory aloneUseful similarity search, but it does not know whether a fact is stale, contradicted, project-specific, or approved.
memblA local Rust daemon stores durable facts, filters by lifecycle status, ranks with SQLite FTS, and logs what was recalled.
live graph
indexFTS5lexical recall for M1filteractivestale facts excludedscopeprojectfacts stay with the repopacket8 maxcompact recall by defaultledgerTTLrepeated recall is measuredreviewhumancandidates promote or rejectindexFTS5lexical recall for M1filteractivestale facts excludedscopeprojectfacts stay with the repopacket8 maxcompact recall by defaultledgerTTLrepeated recall is measuredreviewhumancandidates promote or reject
Rust
local daemon and CLI
SQLite FTS5
current recall index
127.0.0.1
default daemon bind
5 statuses
excluded from recall by default

how it works

The value is smaller, cleaner context.

membl is for teams that run coding agents repeatedly on the same codebase. It reduces the setup tax at the start of a session and makes memory review explicit instead of magical.

01

Capture what should survive the chat

Hooks, imports, and the CLI turn agent sessions into candidate memories instead of another pile of transcript text.

02

Keep state with every fact

A memory is not just a string. It has project scope, kind, confidence, source evidence, and a lifecycle status.

03

Return a small packet

The next agent gets an active, project-scoped recall packet instead of raw notes, stale decisions, or the whole history.

pricing

Run the engine. Or let us run it for you.

The recall engine is open source and free to run yourself. The dashboard, shared graph, and managed operations come with Cloud. That's what you're paying for.

Open Source

CLI · MIT-licensed · self-managed
$0

The engine, on your machine. Terminal-only, no hosted dashboard.

  • Rust recall and conflict-resolution engine
  • CLI: capture, recall, review, savings
  • Claude Code and Codex adapters
  • Local SQLite store, self-managed
  • Community support through GitHub issues
Request access →

Enterprise

self-hosted · 50+ seats · your VPC
Custom

The full Cloud experience, dashboard included, running inside your perimeter.

  • Everything in Cloud, self-hosted
  • SAML, SCIM, custom retention policy
  • Procurement paperwork, MSA, DPA
  • Priority support and a shared Slack channel
Talk to sales →

security & privacy

Same engine, three ways to run it.

The architecture is the same whether you self-manage, use Cloud, or self-host. Six defaults you don't have to configure.

verified

You choose where the engine runs

Open Source on your laptop, our managed Cloud, or self-hosted Enterprise inside your VPC. The recall and conflict logic is identical across all three.

verified

Provenance on every memory

Memories can link back to raw events, session payloads, or imported evidence. The dashboard exposes that evidence instead of hiding it inside a vector.

verified

Never used as training data

Your memories, sessions, and code are not used to train any model. Not ours, not anyone else's.

verified

Conflicts surface, never silently win

When two captures disagree, membl flags the conflict instead of picking. Stale, contradicted, or superseded facts drop out of recall automatically.

verified

Open, portable storage

The M1 store is SQLite with documented tables for events, memories, edges, recall, review, and savings. Backups and exports stay straightforward.

verified

Measured before claimed

Savings are logged as events. Repeated recall is measured directly; broader context savings stay labeled as estimates until benchmarks prove them.