Memory that coding agents actually use.
membl captures durable project facts from every agent session, resolves them into an inspectable graph, and returns only what is true now — in 12 milliseconds, on a Rust engine, with provenance on every memory.
- auth uses bcrypt, cost 12
src/auth/hash.ts · commit a3f12c
- feature flag `recall_v2` is on in prod
config/flags.json · 2d ago
- ingest worker runs on its own queue
infra/workers.tf · review notes
- old retry policy (superseded)
filtered · stale 7d
Stop explaining your codebase every session.
Coding agents are stateless. CLAUDE.md rots. Latest-notes replay drags stale context into every prompt. membl is the layer in between — a graph of reviewed facts the daemon keeps current while your team works.
Project-scoped facts in the next prompt.
Each agent session starts with what is true now — not a transcript replay or another repo's notes.
Conflicts surface, never silently win.
When two captures disagree, membl flags the conflict. Stale and superseded facts drop out automatically.
Every recall and every byte saved is logged.
The savings ledger shows tokens avoided per session, per repo, and per agent — by confidence tier.
Different from a CLAUDE.md. Different from a vector store.
CLAUDE.md is a file you maintain by hand. A vector store returns whatever was close. membl is a captured graph the daemon maintains for you — scoped, reviewed, and inspectable.
Project scope
Every memory is bound to a repo. No cross-project leakage into recall packets.
Lifecycle status
Active, stale, contradicted, superseded, or rejected — set at review, enforced at recall.
Provenance on every fact
Each memory links back to a commit, file, or session payload. Inspect the evidence, not a vector.
Recall trace
Every packet shows what was injected and what was filtered out, and why.
Measured, not implied.
The current fixture returns every active fact, excludes stale and superseded ones, and ships a packet a third the size of latest-notes replay — in twelve milliseconds, p50.
See the benchmarkOpen source engine. Hosted dashboard when you want it.
The engine is MIT-licensed and runs anywhere. Cloud adds the dashboard, the shared graph, and the conflict review your team uses every day. Enterprise puts the same UX inside your VPC.
Open Source
The full engine — daemon, CLI, capture hooks, recall. Local SQLite store. No dashboard.
- Rust daemon, p50 12 ms recall
- CLI: capture, recall, review, savings
- Claude Code and Codex adapters
- Local SQLite store, fully portable
- Project-scoped, lifecycle-aware recall
- Savings ledger via CLI
Cloud
Everything in Open Source, plus the dashboard your team actually uses to review, resolve, and recall.
- Inbox, conflict queue, recall trace UI
- Shared graph scoped per workspace
- SSO, audit log, role-based review
- Signed-token recall over HTTPS
- Managed upgrades and backups
- Savings dashboard across sessions
Enterprise
The full Cloud experience deployed inside your own security boundary. Same UX, same engine.
- Everything in Cloud, self-hosted
- SAML, SCIM, custom retention
- Private network and KMS integration
- Procurement paperwork, MSA, DPA
- Priority support and implementation
security & privacy
Your memory stays yours.
The architecture is the same whether you run it locally, in Cloud, or self-hosted. Tokens are scoped per machine. Rejected and stale facts never enter recall.
Provenance on every memory
Memories link back to the commit, file, or session payload they came from. The dashboard exposes that evidence — never hidden inside a vector.
Your code is never training data
Memories, sessions, and code are not used to train any model. Not ours, not anyone else's. Contractual, not aspirational.
Conflicts surface, never silently win
When two captures disagree, membl flags the conflict. Stale, contradicted, and superseded facts drop out of recall automatically.
You choose where the engine runs
Open Source on your laptop, managed Cloud over HTTPS, or self-hosted Enterprise inside your VPC. Identical engine across all three.
Open, portable storage
The store is SQLite with documented tables for events, memories, edges, recall, review, and savings. Backups and exports are straightforward.
Benchmarked before claimed
p50 12 ms recall. 4/4 active facts returned, zero stale contamination, 201-token packet versus 642 for stale-notes replay.