Skip to content

ENGRAM.md

Latest commit

 

History

History
108 lines (76 loc) · 4.7 KB

ENGRAM.md

File metadata and controls

108 lines (76 loc) · 4.7 KB
Raw

Engram — AI persistent memory in this repository

Repository-local product spec and REQ-* traceability (Makefiles, demos) are described in Spec-driven development.

This repository is prepared to use Engram with Cursor (and other MCP-compatible editors). The configuration is already present so that once you install the Engram CLI, your AI agent can use persistent memory across sessions.

Table of contents

What is Engram?

Engram is an MCP (Model Context Protocol) server that gives AI coding agents (Cursor, Claude Code, etc.) persistent memory. It stores context in a local vault (SQLite) so the agent does not need to re-discover project structure and conventions in every session.

Capabilities include:

  • Remember — Store project notes, decisions, and context.
  • Recall — Retrieve relevant memories using semantic search.
  • Briefing — Get a structured context summary at session start.
  • Consolidate — Distill recent episodes into long-term knowledge.
  • Entities & stats — Inspect the knowledge graph and vault statistics.

All data stays local; no code or secrets are sent to external services except optional embeddings (see Engram documentation).

Repository setup

In the root of this repository you will find:

  • .cursor/mcp.json — MCP configuration that registers the Engram server with Cursor.

Content:

{
  "mcpServers": {
    "engram": {
      "command": "engram",
      "args": ["mcp"]
    }
  }
}

Cursor uses this to start the Engram MCP server when needed (via the engram mcp command). No need to edit this file unless you want to change the server name or arguments.

How to install Engram

You need the Engram CLI (engram) on your machine so Cursor can run engram mcp.

Option 1: npm (recommended)

npm install -g engram-sdk
engram init

engram init will detect your editor, prompt for a Gemini API key (for embeddings and consolidation; free at Google AI Studio), and create your vault. You can skip or complete init; the MCP will work as long as the engram binary is in your PATH.

Option 2: Homebrew

brew tap tstockham96/engram
brew install engram-sdk
engram init

Verify

Ensure engram is available:

engram --help
engram mcp   # Starts the MCP server (stdio); Cursor runs this automatically

How to use

  1. Install the Engram CLI as above and run engram init if you want to set up the vault and Gemini key.
  2. Restart Cursor (or your MCP client) so it loads .cursor/mcp.json.
  3. Use the agent as usual. When the Engram MCP is connected, the agent can use tools such as:
    • engram_remember — Store a memory (episodic, semantic, or procedural).
    • engram_recall — Retrieve relevant memories for the current context.
    • engram_briefing — Get a context summary for the session.
    • engram_consolidate — Run consolidation (distill episodes into knowledge).
    • engram_surface, engram_connect, engram_forget, engram_entities, engram_stats, engram_ingest — See Engram docs for details.

No extra steps are required in this repo: open the project in Cursor, ensure engram is installed and on your PATH, and the agent can use Engram when the MCP server is enabled.

Clearing or resetting the vault

  • Remove a single memory: Use the agent tool engram_forget with the memory id (from recall/stats), or from the CLI: engram forget <id> (soft delete) or engram forget <id> --hard (permanent). You can get ids via engram recall <context> or engram stats (and the agent can use engram_entities / engram_stats).
  • Empty the vault completely: The vault is a local SQLite file. Default location: ~/.engram/default.db (or the path set in ENGRAM_DB). To wipe all memories, delete that file or the whole directory: 
    rm -f ~/.engram/default.db
# or, to remove the whole vault directory:
rm -rf ~/.engram
    After that, the next time you use Engram a new vault will be created. You can run engram init again if you want to reconfigure the Gemini key or editor.

References