Your agents forget.Engram remembers.
One shared memory for Claude Code, Codex, Cursor and the rest of your MCP agents. Save a decision in one, recall it in the next, and stop re-explaining your project every session.
Closed beta. New invites get 60 days of unlimited knowledge transfers, no card. Have an invite?
Works with the agents you already use
Claude Code
Codex
Cursor
ChatGPT
Grok Bot
Muse
InstinctWindsurf
Devin
OpenClaw
Hermes Agent
Gemini CLI
OpenCode
Cline
Goose
Amp
GitHub Copilot
Visual Studio Code
Aider
Grok
Claude
MCP support and setup vary by client. See setup for each one
Decide in Claude Code.Continue in Codex.
One agent saves the decision. The next one recalls it. You never paste the same context twice.
› Remember: background-job retries must use a stable request ID. ⏺ engram · remember ✓ Added a dated entry to projects/billing
› What did we decide about job retries? ⏺ engram · recall "job retries" ✓ 1 snippet from projects/billing “Background-job retries must use a stable request ID.” Got it. I’ll key the retry on the request ID.
Example session. You type the prompts; your agent decides when to call Engram.
Everything your agents need to stop starting over.
Memory you can actually read
Decisions land on readable pages with tags, links and a timeline. Open the console and see exactly what was saved.
How memory worksRecall from any agent
A second agent asks, Engram returns a few relevant snippets. Full pages are one call away.
How recall worksA trail you can trust
Engram’s own calls are observed. Work in other tools appears when an agent reports it, and stays labeled unverified.
What gets recordedShare a slice, not the whole drive
Describe what a team or agent should see. Review the exact excerpts, approve them, and share a fixed snapshot. Everything else stays yours.
How scoped sharing worksOpen source. Run it yourself.
The gateway, memory, activity and console are MIT licensed. Docker Compose is the starting point.
docker compose up --build -dOptional AI stays off
External AI processing starts off. Storage, keyword search and activity work without it until an owner turns it on.
One key per agent
Every connection belongs to one workspace. Turn off write access for agents that should only read.
Know the boundariesConnected in one command.
Guided setup checks identity and tool discovery before it writes your agent’s native config. Your key goes in at a hidden prompt, never into your shell history.
- Claude Code, Codex, Cursor, OpenClaw, Hermes Agent, Amp and Devin CLI in one step
- Claude and ChatGPT connectors sign in with OAuth
doctorchecks a connection without changing it
# Connect every supported coding agent on this machine $ npx -y @ani-hq/engram-mcp@latest
# Native HTTP MCP configuration $ npx @ani-hq/engram-mcp setup claude
# Guided configuration, with OAuth also available $ npx @ani-hq/engram-mcp setup codex
# A named server in Cursor’s MCP configuration $ npx @ani-hq/engram-mcp setup cursor
# Read-only: checks identity and the tool catalog $ npx @ani-hq/engram-mcp doctor codex
Shared with your agents.Private by default.
Scoped to a workspace
Connections belong to one workspace. Personal memory stays separate from a team’s shared space, and you choose what crosses over.
Nothing silently watched
Engram observes calls to its own memory tools. It does not monitor every conversation or tool call your agents make.
Open source, your call
Use hosted Engram or run it yourself under MIT. Self-hosting puts the infrastructure, updates and backups in your hands.
Questions, answered.
Looking for setup help? Start with the support guide
What exactly do my agents share?
The memory pages and activity in the workspace they are authorized to use. Save useful context through Engram, then recall it from another connected agent. A separate workspace has separate memory storage and scoped access.
Does Engram automatically record everything?
No. It observes requests to its own memory tools. External actions appear when an agent or integration reports them, and those reports remain labeled unverified. A connection alone does not capture every tool call or conversation.
Do I have to change the agent I use?
Keep your existing MCP-compatible client. Guided setup supports Claude Code, Codex, Cursor, OpenClaw, Hermes Agent, Amp, Devin CLI, VS Code and Windsurf. Chat clients that need OAuth use their connector sign-in flow. See the connection options.
How do I get an invitation?
Hosted Engram is in closed beta. Ask for an invite on X. Open the single-use invite link and verify your chosen inbox; an email-bound invitation needs its invited address. New friend invitations include 60 days of unlimited knowledge transfers after verification. Jev is separate. Existing members sign in with email.
What lives on this shared drive?
Engram holds memory pages and project context: notes, decisions, and the useful details your agents save through MCP. Task activity appears alongside that memory. The shared-drive analogy describes a common place for agent context; it does not promise general file upload or automatic synchronization of every app.
Can I give an agent only part of my context?
Yes. Create a scoped share in Sharing, review the exact excerpts, and configure its separate read-only agent connection. If the agent already has a broader workspace connection, disconnect or revoke that separately. You can also share approved excerpts with a team; everyone authorized there can read them. See how scoped shares work.
Do scoped shares stay in sync?
A reviewed share is a fixed snapshot. New or edited source text needs fresh review and approval. Revoking a share stops future service reads through it; it cannot recall text someone already read or exported.
Can I run it on my own machine?
Yes. The MIT-licensed app includes the MCP gateway, memory, activity history, and web console. Docker Compose is the local starting point. You operate the infrastructure and any optional providers. Read the self-hosting guide.
Stop re-explaining your project.Give your next agent a head start.
Engram Cloud is in closed beta. Ask for an invite, or run it yourself today.