Memex
Your repo remembers the code. Memex remembers you and your work.
A context layer you own that works with any AI agent: the durable stuff, like preferences, decisions and their reasons, people, plans and background, shared across every tool you use. Code-specific context stays where it already lives (your repo, CLAUDE.md, AGENTS.md); Memex holds what a repo can't, so you don't re-explain yourself in each new app.
Tell your agent: "Set up a Memex for me using this repo" and it follows SETUP.md. Later, tell any agent "sync memex" and it follows SYNC.md.
- Your data, your Supabase. Nothing is hosted by anyone else. Cost is pennies a month.
- Any agent. Memex is an MCP server over HTTP. Tested with Claude and Muse reading and writing the same memory; Codex, Cursor and other MCP clients connect with a URL and a key.
- Sync. Agents push durable facts and pull what the others learned, incrementally, without duplicates, in pages.
- Hands-off, if you want it. Ask any agent that can run scheduled tasks to "sync memex every morning" (pull-only: it fetches what other agents saved and summarizes it) or "sync memex every morning and save what's worth keeping" (pull and save: your yes at setup is the standing approval). Unattended saves are limited by the server itself to a small number of short notes per run (10 by default; configurable), skip anything secret-like, skip repeats of what the agent already saved (or what you retracted), and are tagged
auto-sync. A retract_thoughts tool can hide exactly those notes, reversibly, if one was wrong.
- Search that finds things. Meaning and exact words are searched together and fused. Recent notes can be favored on request. Notes that a later note corrected are flagged, ranked lower, and shown next to their current version. A note that re-states something the user already corrected (easy for an unattended agent to do) is flagged as disputed, and the correction is shown first.
- Dreaming (optional). A nightly job uses a model to tag, summarize and connect memories, resolve relative dates, write a short digest, and, once enough notes pile up about one topic, write an overview note of the current picture (what it is, key decisions, state, open questions). Overviews are linked back to their source notes, use corrections and skip outdated or disputed notes, and are refreshed (the old version marked replaced) when new notes join the topic. It never deletes or rewrites what you wrote, and every change is logged and reversible. It runs only if you add a model API key.
- Careful with secrets. Anything that looks like a credential is flagged
restricted: never sent to a model or an embeddings provider, and hidden from search and sync unless explicitly requested. Agents are told to treat memories as evidence, never as instructions.
Status
Early, and in daily use by its author across several agents. The setup was tested end to end on a brand-new Supabase project (which found and fixed a migration bug); the database migrations are also tested on an empty database and on the old Open Brain table shape (see below). Issues and setup reports are welcome.
Layout
| Path | What it is |
|---|
SETUP.md, SYNC.md | Instructions written for agents to follow |
supabase/migrations | Database schema, 7 idempotent migrations |
supabase/functions/memex-mcp | The MCP server: capture, search, list, stats, sync, retract |
supabase/functions/memex-dream | The nightly Dreaming job |
supabase/functions/_shared | Secret detection and output cleaning used by both |
supabase/schedule-dreaming.sql | Template that schedules Dreaming (needs your project details; never commit it filled in) |
scripts/doctor.mjs | Checks an install end to end; output is safe to paste into a report |
scripts/check-secrets.mjs, .githooks | Blocks commits that contain secrets or personal details |
scripts/generate-key.* | Access key generation |
tests/ | Migration tests on an in-process Postgres |
docs/ | Landing page (GitHub Pages) |
Try it
node scripts/doctor.mjs # checks your install (reads MEMEX_URL and MCP_ACCESS_KEY from .env)
node scripts/check-sync.mjs # reports on what scheduled syncs saved (counts only; --show adds previews)
cd tests && npm install && npm test # runs all migrations on an empty database and on the old schema
git config core.hooksPath .githooks # turn on the secret scan before each commit
What leaves your project
- Embeddings: each memory's text goes to the embeddings provider (OpenRouter by default) so it can be searched by meaning. Restricted memories are skipped.
- Dreaming: if you enable it, new memories go to the model provider you configure (Meta first, with an optional fallback). Restricted memories are skipped. Leave the key unset and nothing is sent.
Credits
Building on the excellent Open Brain project by Nate B. Jones, and on an earlier write-up, Liberating memory. No code from either is used here; the ideas are credited, and the implementation is new.
The logo shapes (bookmark and sparkle) come from Lucide, ISC license.
