Engram MCP Server - Git-Backed Memory Dashboard¶

Engram is a self-hosted MCP server + dashboard that gives AI agents shared memory over a git-backed folder of markdown files. What differentiates it from other memory servers: search is authoritative - it always returns the latest committed version, not a stale embedding. The dashboard provides human visibility into what agents are remembering.

Source: awesome-mcp-servers PR #9807 (discovered July 19, 2026) Category: Knowledge & Memory Author: rwnalds Repo: https://github.com/rwnalds/engram Status: Active

Why This Matters¶

Most vector-based memory solutions have a fundamental problem: the embedding and the source document drift apart. Engram solves this by making the source of truth a git repository of markdown files - every write is a commit, every read is from HEAD. The dashboard lets humans audit what agents are writing, edit memories directly, and roll back bad writes via git.

Installation¶

Engram is a self-hosted server with a persistent volume, so run it as a container - Railway, Render, Fly, or your own box. Serverless (Vercel) will not work: Engram holds a volume, a file watcher and an in-memory index that a serverless function cannot keep alive.

# Prebuilt image (the -app image serves HTTP; plain `engram:latest` is the stdio server)
docker run -d \
  -p 3000:3000 \
  -v engram-data:/data \
  -e ENGRAM_DATA_DIR=/data \
  ghcr.io/rwnalds/engram-app:latest

Then open the dashboard, connect your vault repo under Workspaces (by URL + token, or GitHub OAuth), and create an MCP token on the Connect page.

Note on the package name: npm install -g engram-mcp is a different, unrelated product (ngrm.ai). It is not this server and does not expose these tools.

Connecting an MCP Client¶

Engram speaks HTTP at POST /api/mcp with a bearer token. Token scope decides what the agent sees: a read token never sees the write tools.

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

Key Tools¶

15 tools, split by token scope. brain_capture appears only when the Curator is set to full.

Tool	Access	Description
`brain_search`	read	Search the active vault
`brain_read`	read	Read a note
`brain_list`	read	List notes in a folder
`brain_recent`	read	Recently changed notes
`brain_tree`	read	Folder tree for the vault
`brain_backlinks`	read	Inbound links to a note
`brain_graph`	read	Link graph around a note
`brain_schema`	read	Vault schema and conventions
`brain_write`	write	Create a note
`brain_edit`	write	Edit an existing note
`brain_append`	write	Append to a note
`brain_move`	write	Move or re-file a note
`brain_supersede`	write	Replace a note, keeping the prior version
`brain_create_folder`	write	Create a folder
`brain_delete`	write	Delete a note

The read/write split is enforced server-side: an agent holding a read-scope token cannot call the write tools at all.

Dashboard¶

Engram includes a web dashboard at http://localhost:3001 where humans can: - Browse all agent-written memories - Edit or delete entries manually - View git history and diffs - Roll back to any previous version

Integration with CorpusIQ¶

Engram's git-backed approach is philosophically aligned with how CorpusIQ already operates (GBrain's markdown vault, git-tracked docs). Could serve as a lighter alternative to GBrain for teams that want agent memory without the full knowledge graph infrastructure.