Shared memory for AI coding agents. Every agent session leaves a layer. The next one reads it.
What one agent learns, the next one knows. Decisions, conventions and failed attempts saved in Cursor are found by Claude Code, Codex or ChatGPT, for every developer on the project. A hosted MCP server, with review, roles and history.
14 days of Team, no card needed. Set up in about five minutes.
- today Refund flow done. Next: receipt emails. Handoff, Sam in Claude Code
- 2 days memory_limit did not fix the import timeout Failed attempt, Priya in Cursor
- 5 days UUID primary keys on every new table Decision, Priya in Cursor
- 2 weeks Feature tests run against Postgres, never SQLite Convention, Sam in Codex
- 1 month Staging resets every Sunday at 02:00 UTC Fact, Ana in ChatGPT
- 3 months Payments go through Dodo Decision, Ana in Claude.ai
Let's use UUID primary keys on every new table. Integer ids leaked order counts through the invoice API.
Called memory_save
- type
- decision
- title
- Use UUID primary keys
- status
- approved
Saved to the project memory. Every agent on this repository will find it.
> add an invoices_archive table
● retain7: memory_search
query: "primary key type for new tables"
└ 1. Use UUID primary keys decision, approved, 2 days ago
● Following the team decision on primary keys.
Created database/migrations/create_invoices_archive_table.php
$table->uuid('id')->primary();
>
How it works
Set it up once. Every session after that starts informed.
-
Connect your tools
One command for Claude Code, Cursor, Codex or OpenCode. Claude.ai and ChatGPT connect with a sign-in.
claude mcp add --transport http --scope user retain7 https://retain7.com/mcp
-
Agents save what they learn
When you make a decision, set a convention or watch an approach fail, the agent saves one short, typed memory. Secrets are refused before anything is stored.
-
The next agent picks up
Sessions open with the last handoff and the memories that matter most. Agents search before they answer, by meaning and keywords, so their own words find the right memory.
Control
Your team decides what agents remember
- Review inbox
- New agent memories wait for an owner or admin to approve them. Turn on auto-approve per project when you trust the flow.
- Outdated memories
- When code contradicts a memory, agents mark it outdated with the reason. It ranks lower and carries the warning.
- Roles
- Owners, admins, members and read-only viewers. Viewers' agents can search but never save.
- Handoffs
- Agents leave a note when they stop: what was done, what is next, what failed. The next session starts from it.
- Full history
- Every edit is versioned, with who made it and from which tool. Restore any version in one click.
- Yours to take
- Export any project as JSON or Markdown. Delete a project, team or account and it is gone within a day.
Security
Safe for real codebases
Agent memory sits close to your code, so it is built to keep secrets out and teams apart.
-
Secrets are refused
API keys, tokens, private keys, passwords and credentialed connection strings are blocked on every save path.
-
Teams stay separate
Every lookup goes through a project the person can access. Automated tests cover each team boundary.
-
Standard sign-in
OAuth 2.1 with PKCE for Claude.ai and ChatGPT, revocable personal tokens for local tools.
-
No third-party AI
Search vectors come from a model on our own servers. Your memory text is not sent to an AI provider by us.
Pricing
Simple pricing
Flat prices. No charge per memory or per search. Every account starts with a 14-day Team trial.
Free
Try it on one repository.
$0 forever
No card needed.
Get started- 1 member
- 1 project
- 300 memories
- 200 agent searches a day
- Every MCP client, hooks and handoffs
Pro
For one developer with many repositories.
$9 per month
Or $90 a year.
Get started- 1 member
- Unlimited projects
- 10,000 memories
- Unlimited agent searches
- Every MCP client, hooks and handoffs
Team
RecommendedFor agencies and teams sharing projects.
$8 per member a month
Minimum 3 members. 14-day free trial with up to 5 members and 300 memories.
Start free trial- Unlimited members
- Unlimited projects
- 50,000 memories
- Unlimited agent searches
- Every MCP client, hooks and handoffs
- Roles, review inbox and activity log
Questions
Where is my memory stored?
In our Postgres database. Search vectors come from an embedding model we run ourselves, so your memory text is not sent to a third-party AI API.
Will an agent follow instructions hidden in a memory?
Retain7 tells every agent that memory content is data written by people and other agents, not instructions, and every briefing ends with that reminder.
Does it read my code?
No. Agents save only what they choose to file through the memory tools. The optional Claude Code hooks send the last assistant message of a session to write the handoff.
What happens when the trial ends?
The team moves to the Free plan unless an owner subscribes. Nothing is deleted. Saving stops once you pass the Free limits.
Can I use it alone?
Yes. Pro is for one developer who switches between tools and repositories.
Which tools work?
Any MCP client that connects over HTTP. The docs have step-by-step setup for Claude Code, Cursor, Codex, Gemini CLI, Pi, Hermes Agent, Zed, Claude.ai, ChatGPT and more.
Stop re-explaining your project to every agent.
Free for 14 days on the Team plan. No card needed.