One memory for every AI agent in your company. It runs in your AWS account.
Claude, ChatGPT, Codex and any MCP client search the same runbooks and code, with a key and a daily model-spend cap for each person. MIT licensed. Pilot it in two to four weeks.
- Open source, MIT
- No vendor analytics or third-party telemetry
Your knowledge, connected.
From one useful note to a shared memory. Explore what Memrain can do today.
Agents
We configure specialist agents with the right roles, tools, skills and shared Memrain context. From a personal assistant to a coordinated team, each agent has a clear task and a defined way to work.
Use cases
Customer context, meetings, onboarding and knowledge migration. See how teams bring memory and the tools of their connected agent into the same work.
Workspace
Keep personal context and team knowledge organized in infrastructure you control.
Pages
Capture notes, decisions and source material as linked, versioned pages your assistants can find again.
Memory
Find the decision, the person and the reason behind the work. Bring the relevant context into your next conversation.
Tools
Search, recall, update and inspect knowledge through one Model Context Protocol connection.
Skills
Inspectable instructions help your connected assistant capture, research and maintain useful knowledge. You choose the work and the tools it may use.
Works with
Connect the assistants you already use to one memory. Choose a client guide and check its connection requirements.
Imports
Bring supported conversation exports, notes and code into a searchable memory. Import deliberately, with source context intact.
One store your agents share, in infrastructure you already govern.
-
Same memory, whichever vendor
People on Claude and people on ChatGPT search the same runbooks and code. A shared team source sits in each person's read set, so switching or mixing model vendors does not split what the team knows.
Existing Claude, ChatGPT and Codex history can be imported, redacted on the way in.
-
Your account, your rules
The Terraform stack runs in your AWS account. Stored data stays there. Passages your agent retrieves go to that agent's model, and with the default Cloudflare Tunnel requests pass through Cloudflare. Caddy mode takes Cloudflare out of the path.
How data flows -
A key per person
Each person gets their own credential, and you can revoke one person without touching anyone else. Daily USD caps apply per person, per connector or deployment-wide, on priced model calls.
A private space per person is available. It is pilot-grade, and its known limits are documented.
See what your agents ask and what it costs.
An admin panel with its own sign-in shows 7, 30 and 90-day activity, credentials and members, spend and a request log. Caps apply to priced model calls, and some costs are approximations.
Results come back with the page they came from.
Memrain is the memory behind your agent, not a chat window. Your agent writes the answer.
- Claude
- ChatGPT
- Codex
/mcp address- One EC2 instance
- RDS Postgres 16 with pgvector
- Embeddings on Amazon Bedrock
- Secrets Manager
Load your material
Runbooks, notes, decisions, repositories and chat exports. Credential-shaped text is redacted before it is stored.
Memrain indexes it
Hybrid search over meaning and exact words, a code graph for TypeScript, Python and Go, and facts and timelines between pages.
Your agent asks over MCP
Each passage comes back with a pointer to its source page, so anyone can open it and check.
- Storage
- RDS Postgres 16 with pgvector
- Embeddings
- Amazon Titan v2 on Bedrock
- Interface
- 96 MCP tools, OAuth 2.1
- Ingress
- Cloudflare Tunnel or Caddy
- History
- Every page versioned, revertible
- Export
- Pages to markdown
Also works for one person.
You do not need a team. Connect Claude or Codex to Memrain and keep notes, decisions and code in one memory you can search. It costs about $52 a month on AWS for the server and database, plus model usage.
Read the quick startAsk from
- Claude
- ChatGPT
- Codex
Remember from
- Markdown notes
- Code repositories
- ChatGPT and Claude exports
- Claude and Codex coding sessions
# fork the repository, then
git clone https://github.com/<you>/memrain.git
cd memrain
make init # a few questions, writes your config
make plan && make apply
# connect your agent
claude mcp add --transport http memrain https://<your-host>/mcpShort answers
Why not Claude Projects or ChatGPT memory?
Those keep memory per vendor, in the vendor's cloud. Memrain is one store that every MCP client reads, in your AWS account, with a key per person.
Does data leave our account?
Stored data does not. What your agent retrieves goes to that agent's model, and in the default Tunnel mode requests pass through Cloudflare. Caddy mode removes Cloudflare.
Do you collect telemetry?
There are no vendor analytics and no third-party telemetry. Memrain keeps its own usage counters and logs, inside your account.
Is there single sign-on?
No. People connect through your organisation's OAuth connector with a one-time enrollment code each. There is no SAML, no OpenID Connect federation and no SCIM.
Can we run it for real?
It is one EC2 instance and an RDS database with 7-day automated backups, encrypted storage and deletion protection. High availability is not offered. Run a restore drill during your pilot; RTO and RPO figures are not published.
What does it cost?
The software is free under the MIT licence. Running it costs about $52 a month on AWS for the server and database, plus model usage. That figure is approximate, for eu-west-1, as of October 2026.
Who supports it?
GitHub issues, and e-mail to me if you want help with a pilot. There is no support contract or SLA.
Pilot it in your own AWS account.
I help you stand up one stack, enroll your first people and run a restore drill. No contract, no price.