Engram

The second brain your AI agents read and write.

_{▶ Watch the full-length video (real-time, full quality)}

↗ Try the live demo — no signup. Search retainer price and watch the retired note get withheld.

Engram is a self-hosted MCP server + dashboard that gives Claude Code, Cursor, Hermes, and any Model Context Protocol agent shared, long-term memory they read and write — over a plain, git-backed folder of markdown you own. Built for the case a single agent's memory never hits: a team running several agents against one brain.

Autonomous agents forget everything between sessions — and worse, they can't tell what they remember is still true. An agent pulls an old README, a retired price, an API doc you changed months ago, and quotes it with full confidence, because keyword and vector search both rank by resemblance, not truth. Engram makes "is this still true" a first-class, written property: mark a fact superseded or expired and search withholds it — and tells the agent what it skipped and why. Per-agent read/write tokens and a git audit trail of who-wrote-what keep it sane when the writers are a fleet, not just you.

Unlike a headless memory store, you can watch it happen. A fast dashboard lets you search your brain, see exactly what every agent and teammate changed (with per-file diffs), jump back into recent notes, and curate it all — while agents read and write the same vault over one MCP endpoint. No database: your .md files are the source of truth, git is the durable store, and an in-memory index powers full-text search + a wikilink knowledge graph.

Opinionated about how it stores memory — git-backed markdown, no database, agents write (not just read), self-hosted. Unopinionated about what you keep in it — any markdown vault, any folder structure, any MCP client. Point it at a fresh repo or your existing Obsidian vault: no import step, no lock-in.

What it's for · How it compares · Features · Works with · Quick start · MCP tools · Deploy · FAQ · Contributing

What it's for

Shared memory for a team running multiple agents — one vault, many agents reading and writing concurrently, with per-agent read/write tokens and a git audit trail of who changed what.
Memory that knows what's still true — retire a price, a term, or a changed API doc and your agents stop quoting it; they're told what they skipped and why. The failure a single agent's memory never fixes.
Long-term memory for Claude Code and other coding agents — stop re-explaining your project every session.
A self-hosted, Obsidian-compatible second brain exposed over MCP — your notes, your server, your git repo.
Memory you can see, not a black box — a dashboard to search, watch (with diffs), and curate what your agents remember.
Markdown RAG without the vector database — full-text search + a link graph over human-readable files.

How it compares

Most agent memory is built to answer "what did I store about this?" Engram is built to answer "what is still true about this?" — a different question, and the one that bites when an agent quotes a price you retired months ago.

	Engram	Vector-store memory _{(mem0, Zep, …)}	Markdown memory _{(Basic Memory, …)}	Plain RAG over docs
Storage	markdown files, git is the database	embeddings in a vector DB	markdown files	embeddings in a vector DB
Ranking	relevance × authority	similarity	relevance	similarity
Knows a fact is retired	✅ `superseded_by` + `valid_until`, enforced at search	—	—	—
Explains what it withheld	✅ `excluded[]` with a reason per note	—	—	—
Retire + replace atomically	✅ `brain_supersede`, one commit	—	—	—
Refuses contradicting writes	✅ at write time, not read time	—	—	n/a (read-only)
Audit trail	✅ git, per-write attribution, per-file diffs	varies	git, if you commit	—
Per-agent access control	✅ read / write token scopes	varies	—	—
Human UI	✅ dashboard, search, diffs, graph	varies	—	—
Runs on	your box, one container	mostly hosted SaaS	your box	your box

The row that matters is the third one. Similarity search cannot tell a contradiction from a duplicate — a retired price and a live one are textually identical, so the retired one often outranks the live one by being longer and more detailed. That can't be fixed at read time, which is why Engram writes the retirement down when it happens.

Engram dashboard showing a note marked locked, superseding an earlier pricing note, with backlinks

_{This exact pair ships in sample-vault/. Run bun dev, search acme price,
and watch the retired note get withheld with a reason.}

_{Categories, not feature-by-feature audits of specific products, and accurate to the best of my
knowledge as of July 2026. If something here misrepresents a tool you maintain, open a PR — I'll fix it.}

Features

MCP server — 15 brain_* tools over one bearer-authenticated HTTP endpoint (POST /api/mcp, streamable HTTP JSON-RPC). Connect any MCP client to a single URL. Per-agent token scopes: a read-only token never even sees the write tools.
Human dashboard — a search-first home, file tree, note viewer with Obsidian callouts, wikilinks, and backlinks, Preview / Edit / Split editor with autosave, ⌘K search + in-page keyboard navigation, "jump back in" recents, and a force-directed knowledge graph.
Authority-aware search — ranking knows relevance, not truth, so a superseded note repeats your query words as often as the live one. Every hit carries an authority (authoritative → current → provisional → superseded → archived) derived from the note's folder and frontmatter — so your agents quote the locked doc, not the dead one. Markdown RAG that won't hand back yesterday's answer.
Temporal validity + explainable rejection — mark a fact superseded_by another note or give it a valid_until date, and search withholds it by default (even if it's locked) — then hands the agent an excluded list of what it skipped, each with a reason ("expired 2026-06-01"). One atomic brain_supersede retires the old fact and links the new one in a single commit, so add-and-retire can't drift apart. This is the difference between an agent that remembers and one that knows what's still true.
Write-time contradiction guards — authority ranking fixes reading; these stop the vault accepting the contradiction in the first place. Engram refuses to create a second live note on a subject a live note already covers (the acme-pricing-2026.md-beside-acme-pricing.md bug) and points the agent at brain_supersede instead; refuses to overwrite a note the caller hasn't read; and warns when a status: isn't a word the ranking model knows, so a typo can't silently strip a note's authority.
Audit trail + access control — every write is attributed in git to the token or human that made it, with expandable per-file diffs in the activity feed. Give an agent a read-only token and it never even sees the write tools; a write token can create, edit, move, and archive.
The Curator (optional) — Engram's built-in agent harness over your vault. Chat with your notes (grounded answers, wikilink citations). Or hand brain_capture a rough dump — a meeting note, a voice transcript — and an agentic loop searches what already exists, then files, merges, or archives and returns a manifest of what it touched. It reads before it overwrites and never deletes. Opus / Sonnet / Haiku, your key.
Markdown-native — plain .md + YAML frontmatter + [[wikilinks]]. Drop in an existing Obsidian vault and it just works.
Git-backed — optional auto commit + push of every change. Full history, no lock-in, your data lives in your repo.
No database — files are the source of truth; an in-memory MiniSearch index + a ported wikilink graph power search and backlinks. Nothing to provision.
Multi-workspace — connect multiple vault repos (URL + token or GitHub OAuth), rename, switch the active one, or remove them — all from the UI.
Self-hosted — one Docker container. Railway / Render / Fly / any host with a volume. Not serverless (it needs a persistent volume, a file watcher, and a long-running index).
Team auth — Google SSO + email allowlist for the dashboard; per-agent bearer tokens — or OAuth for Claude.ai custom connectors — for MCP, created/revoked in the UI. Secrets encrypted at rest.
Runtime config — toggle git-sync and the Curator right from the home; manage commit author, keys, and OAuth in Settings — no redeploy.

Works with

Any client that speaks the Model Context Protocol — one endpoint, bearer-token auth. Most-used first:

Claude Code — Anthropic's agentic coding CLI
Codex — OpenAI's coding agent (CLI + IDE)
Hermes — always-on autonomous agent runtime
openclaw — open-source coding agent
Cursor — AI code editor
Cline — VS Code agent
Windsurf — agentic IDE
Claude Desktop — Anthropic's desktop app
…and any other MCP client — Continue, Goose, Zed, Amp, and the rest

If it speaks MCP, it can read and write Engram as shared memory.

Quick start

Want to poke at it first? There's a live demo — open, no signup, resets a few times a day. Break it all you like.

bash

bun install
bun dev            # http://localhost:3000 — runs against ./sample-vault

Point it at your own vault:

bash

VAULT_DIR=/path/to/your/obsidian-or-markdown/vault bun dev

Two ways to run it

Hosted mode (team): the dashboard + HTTP MCP server above — self-host it once, many agents and teammates connect over POST /api/mcp. This is the main mode.
Local mode (stdio): a plain stdio MCP server over a folder, no HTTP/auth/git — for a single machine, Claude Desktop / Cursor, or a registry's Docker introspection:
bash
```
bun run mcp:stdio /path/to/your/vault      # defaults to ./sample-vault
```
Same brain_* tools. Built from Dockerfile.mcp.

Each mode ships as its own image, and they are not interchangeable:

Image	Mode	Use it for
`ghcr.io/rwnalds/engram-app`	HTTP — dashboard + `/api/mcp`	Railway, Render, Fly, any host. This is the one you deploy.
`ghcr.io/rwnalds/engram`	stdio — JSON-RPC on stdin/stdout	Claude Desktop, Cursor, MCP registries. Never opens a port.

Deploying the stdio image as a web service is the one mistake worth calling out: it can only 502, because there is nothing listening. It now refuses to start on a PaaS and tells you this instead.

MCP tools

Agents only ever see the active vault — no repo, workspace, or GitHub tools are exposed. A read-scope token sees only the read tools. brain_capture appears only when the Curator is full.

	Tools
Read	`brain_search` · `brain_read` · `brain_list` · `brain_recent` · `brain_tree` · `brain_backlinks` · `brain_graph` · `brain_schema`
Write (needs a `write`-scope token)	`brain_write` · `brain_edit` · `brain_append` · `brain_move` · `brain_supersede` · `brain_create_folder` · `brain_delete`

Connect an agent (the dashboard → Connect page shows the exact command + token):

Terminal

claude mcp add --transport http engram https://<host>/api/mcp \
  --header "Authorization: Bearer <token>"

Deploy

Runs anywhere you can run a Docker container with a persistent volume — Railway, Render, Fly, or your own box. Serverless (Vercel) won't work: Engram holds a volume, a file watcher, and an in-memory index that a serverless function can't keep alive.

Deploy this repo (root Dockerfile), mount a volume at /data, set ENGRAM_DATA_DIR=/data.
Connect your vault repo(s) in the dashboard (Workspaces) — by URL + token, or GitHub OAuth.
Sign in with Google, create MCP tokens on the Connect page, point your agents at the URL.

Most runtime config (git-sync, AI capture, GitHub OAuth, app name) is editable in the Settings page — only auth/infra bootstrap vars live on the host. Full setup: DEPLOY.md.

Railway: New Project → Deploy from GitHub repo → add a Volume at /data.
Render: one-click via the bundled render.yaml (Docker + a /data disk).

Deploying from a prebuilt image instead of the repo? Use ghcr.io/rwnalds/engram-app:latest — not engram:latest, which is the stdio server and cannot answer HTTP. Set the healthcheck to /api/health.

FAQ

How do I give Claude Code long-term memory? Deploy Engram, connect a markdown vault, and claude mcp add the endpoint. The brain_* tools let Claude Code search, read, and write persistent notes across sessions.

Can multiple AI agents share one knowledge base? Yes. Every agent points at the same MCP URL and reads/writes the same active vault — that's the point. Give each agent its own bearer token, read or write — a read-only token can't mutate your notes.

How do I stop an agent from quoting outdated facts? Retire the fact and search stops surfacing it. When a value changes, call brain_supersede(old, new) — one atomic commit marks the old note superseded_by the new one, and it's withheld from search by default (even if it's locked). Or set valid_until: 2026-12-31 on a note and it self-expires. Retired matches don't vanish silently: brain_search returns them in an excluded list with a reason ("expired 2026-06-01"), so the agent can say what it ignored and why instead of quoting it.

What stops an agent just adding a second, contradicting note? Engram refuses the write. Told "the price is now X", an agent that can't overwrite a note it never read will happily add acme-pricing-2026.md next to acme-pricing.md — nothing is corrupted, and you now have two live notes disagreeing about one number. That write is rejected with a pointer to brain_supersede, which retires the old note and adds the new one in a single commit. Pass allow_conflict: true when both notes genuinely belong.

How do I know what an agent changed? Every write is committed to git attributed to the token or human behind it, and the dashboard's activity feed shows per-file diffs — a built-in audit trail for autonomous agents.

Does it work with my Obsidian vault? Yes. It reads plain markdown with frontmatter and [[wikilinks]], and renders Obsidian-style callouts and backlinks. No import step.

Do I need a vector database? No. Engram uses full-text search (MiniSearch) plus a wikilink graph over human-readable markdown — no embeddings service, no vector store to run.

Can I chat with my notes? Yes — enable the optional Curator, a chat agent that searches and reads your vault to answer with wikilink citations (Opus / Sonnet / Haiku). It's read-only in chat, so it helps you think without changing anything, and it runs on your own Anthropic key.

Can I see what my agents changed? Yes — the Activity view reads your vault's git history and shows every change (agents and teammates alike), expandable to per-file diffs. Since it's just git, you get the full audit trail for free.

Is my data locked in? No. It's just .md files in a git repo you own. Turn Engram off and you still have every note and its full history.

Where does it run / is it self-hosted? You host it. One Docker container on Railway / Render / Fly / any VM with a volume. Your keys, your data.

Contributing

Issues and PRs welcome — especially where the validity model breaks against a vault shaped differently from mine.

CONTRIBUTING.md — setup, conventions, and the pre-PR checklist.
docs/curator.md — how the optional Curator agent loop works.
SECURITY.md — please report vulnerabilities privately, not as an issue.

bash

bun install && bun dev
bun test          # the ranking/authority suite, incl. the stale-truth fixtures

If Engram is useful to you, starring the repo genuinely helps other people find it.

Stack

Next.js 16 (App Router) · React 19 · TypeScript · Tailwind v4 · shadcn/ui · bun · MiniSearch · d3-force · MCP SDK. MIT licensed.

_{Keywords: MCP server · Model Context Protocol · second brain for AI agents · agent memory ·
long-term memory for Claude Code · shared memory for AI agents · self-hosted knowledge base ·
Obsidian-compatible · markdown · knowledge graph · wikilinks · PKM · Zettelkasten · git-backed notes ·
Hermes agent memory · Cursor memory · RAG without a vector database · chat with your markdown notes ·
git-backed agent activity feed · audit trail for AI agents · authority-aware search · read-only vs
write MCP tokens · agent access control · self-organizing notes · agentic note capture · AI that files
your notes · temporal validity · stale memory · agents quoting outdated facts · supersede · note expiry ·
shared memory for a team of agents · Basic Memory alternative · mem0 alternative.}