What you get
- A working governed-memory setup for one team, on your hardware.
- A weekly office hour with the founder for the length of the pilot.
- A findings memo at the end: what worked, what didn't, what governance caught.
Design-partner program
Sovereign Memory Pilot
Give one team's agents a governed, citable, local memory — in a 4–6 week pilot, on your machines, with zero data leaving your building.
MIT open-source, preview stage
The pilot
What it is
MemexLab is the governed memory layer for AI agents — MIT
open-source, local-first, preview stage. In a pilot, we take one
team's (or one executive's) knowledge base, bootstrap it into a
plain-markdown vault on your hardware, and connect your agents
(Claude Code, Claude Desktop, Cursor) to it through
memexlab-mcp:
- Citable answers — deterministic search; every answer points at your actual documents.
- Governed writes — agents can file new notes only into
inbox/, with provenance, every write logged. The canonical layer cannot be modified, by construction. - Sovereign — everything runs on your machines. Nothing reaches us. There is no cloud, no account, no telemetry.
The fit
Who it's for
Regulated and confidentiality-bound teams — finance, legal, healthcare — that want agent leverage without putting their knowledge in someone else's cloud.
The terms
What you get, what we ask
What we ask
- A short written feedback form every two weeks (5 questions).
- Permission to name you, privately, as a reference.
- If the pilot succeeds: an anonymized case study and a letter of intent — public logo use only if you volunteer it.
Architecture
Architecture & security
Where your data lives: on your machines, as plain markdown files you can read, diff, and delete. The MCP server is a local process; agents talk to it over stdio. No cloud service, no account, no telemetry, no network calls from the memory layer.
Write boundary (enforced in code, not policy):
agents can create notes only in inbox/, stamped with
provenance (who, when, cited sources). Every write appends to an
audit log (.memexlab/log.jsonl). There is no code path
that modifies your canonical notes — verified by tests, including
adversarial ones (path traversal, symlinks, injection).
Citations: retrieval is deterministic (BM25, no embeddings); results carry slugs, so answers cite your actual documents — auditable by anyone.
Compliance posture (stated honestly): local-first
collapses most of the GDPR surface structurally — no cross-border
transfer, no third-party processor; you remain sole controller.
SOC 2 Type II is on our roadmap, not a current claim. The engine is
MIT open source
(github.com/btekmen/memexlab-engine)
— your security team can read every line.
Model choice: bring your own — a local model (fully offline) or your own API key. The memory layer itself never talks to any model provider.
Design-partner program
Apply — we run at most two pilots at a time
Pilots are free. We run at most two pilots at a time — apply below and we'll schedule honestly.