Persistent engineering memory

Your coding agents should remember.

One engineering memory for Claude Code, Codex, Grok and Antigravity CLI that belongs to you, your project and your team. Agents check it before they work, save lessons as they go, and get back only what matters for the task.

Run in your terminal
npx memroot@latest setup

Run setup. Your browser opens for sign-in or registration.

memroot · demo1:08

Get connected

Quick start.

Choose your agent. Follow three steps.

memroot · setupQuick start
  1. Run in your repository

    Setup Claude Code
    npx memroot@latest setup --agent claude
  2. Authorize in Claude Code

    Inside Claude Code
    /mcp

    Select memroot, choose Authenticate, then restart Claude Code.

  3. Optional: import existing agent memory.

    Import what your agents already remember
    npx memroot@latest import

01 / Product

Knowledge that takes root.

Engineering lessons belong in a typed memory graph, with evidence, provenance, and Git context.

  • Durable engineering lessons such as decisions, proven fixes, conventions, and failed approaches.
  • Memroot’s hooks prompt agents to check Memroot before each task and to save lessons as they work, without being asked.
  • Advanced, context-aware ranking reads the task and the files in play, and retrieves a small memory pack of only what matters instead of an unbounded history dump.
  • Agents can archive a wrong or outdated memory. It stops being used but is not deleted, and you can restore it from your console.
  • Memroot keeps engineering knowledge independent of any one coding agent.
  • Teams are free. Create a team, invite people by email as owners, admins or members, and choose per project whether each person can read or read and save memories. You can also share a single project with a guest, and every memory shows who saved it.
Example memoryProven fix

Example: account deletion must purge the search index.

Deleting an account must remove its records from search as well as the primary database.

Evidence
src/accounts/delete.ts · test passes
Source
Claude Code session· confirmed by Codex
Git
main @ 4f2a91c
State
Current· gaining relevance

02 / How it works

From session to signal.

  1. Connect your agents.

    One setup command connects Claude Code, Codex and Grok. Memroot’s hooks remind each agent to check Memroot before it works. Browser authorization controls what each connection can access.

  2. Capture the lessons.

    Agents save lessons as they work. Session capture runs during the work and when a session ends, redacts it on your machine and turns it into typed lessons with evidence; the transcript is never uploaded.

  3. Let it dream.

    Dreaming runs in the background, like sleep consolidating the day. It merges memories that state the same fact; you can undo a merge until any memory in it changes. Newer lessons replace the ones they correct, and agents’ feedback shifts relevance.

  4. Recall what matters.

    Memroot’s advanced, context-aware ranking reads the task, the files in play and your Git state, and gives the agent only the memories that matter. When nothing fits, it can return none. Every agent and teammate shares the same memory, on any machine.

03 / Trust

Memory you can trust.

Useful enough to return. Careful enough to deserve a place in your workflow.

Memories belong to the developer, project or team, not to one coding agent.

Review and revoke each coding-agent connection from your console.

Memroot stores structured engineering memories, not full coding-session transcripts.

The memory graph records evidence, provenance, and explicit stale or superseded states.

04 / Pricing

Choose the plan that fits your work.

Free

Available

$0 / month

Limits

  • 1 project
  • 500 memory writes / month
  • 1,000 retrievals / month
  • 2,500 total memory writes
  • 5,000 total assertions
  • 30 data requests / minute

Account registration

Start free

Basic

Available

$9.99 / month

Limits

  • 5 projects
  • 5,000 memory writes / month
  • 10,000 retrievals / month
  • 25,000 total memory writes
  • 75,000 total assertions
  • 60 data requests / minute

Monthly subscription in USD. Review your plan and complete secure checkout after signing in.

Sign in to continue

Choose Basic

Max

Available

$49.99 / month

Limits

  • 100 projects
  • 75,000 memory writes / month
  • 200,000 retrievals / month
  • 375,000 total memory writes
  • 1,125,000 total assertions
  • 300 data requests / minute

Monthly subscription in USD. Review your plan and complete secure checkout after signing in.

Sign in to continue

Choose Max

Limits apply across your tenant. Monthly writes and retrievals reset each UTC calendar month. Total memory writes and total assertions are lifetime admissions; deleting memories does not restore them. An assertion is one atomic statement inside a memory; a memory can hold up to 8.

Teams are free on every plan and have no separate price. A team's seats, which count its people and pending invitations, follow the best plan held by one of its owners: 5 on Free, 10 on Basic, 25 on Pro and 100 on Max.

Works with

One memory. Any agent.

Memroot is built on agent-neutral contracts.

Memory packs give agents relevant context on demand.

Memroot
  • Claude Code
  • Codex
  • Grok
  • Antigravity CLI
  • Teammates

More agents and adapters are on the way.

05 / Questions

The useful details.

What does Memroot remember?

Durable engineering lessons such as decisions, proven fixes, conventions, and failed approaches.

How is Memroot different from CLAUDE.md, AGENTS.md or my agent's own memory?

Memroot works alongside native memory, adding shared project lessons, evidence, correction history and task-specific recall.

Do agents use Memroot on their own?

Yes. Memroot’s hooks remind each agent to check Memroot before it works and to save lessons as it goes, and session capture turns each session into typed lessons on your machine.

How does Memroot decide which memories my agent sees?

Memroot’s advanced, context-aware ranking reads the task, the files in play and your Git state, and returns a small memory pack instead of a whole memory file. When nothing fits, the pack can be empty.

What leaves my machine?

Saved lessons, plus the task, file paths and Git state used for ranking. Session transcripts are redacted on your machine and never uploaded to Memroot, and session capture can be turned off.

How does Memroot keep memory accurate over time?

Memroot dreams: in the background, on a schedule, it compares similar memories and merges duplicates of the same fact into one. The originals are kept, and until the merged memory or one of its originals changes, you can undo the merge in the console. In the console graph, dreaming can also flag two memories that contradict each other. A newer lesson replaces the one it corrects, with the history kept. Agents archive outdated memories instead of deleting them. Their feedback adjusts each memory’s relevance, and you can review everything in the console.

Can my team share a project's memory?

Yes. Create a team in the console, invite people by email and give each member read or read-and-save access to the projects they need. When someone connects a coding agent they choose a personal or team workspace, and the agent works with that person's access only.

Which coding agents work with Memroot?

Memroot connects to Claude Code, Codex, Grok and Antigravity CLI. More agents and adapters are on the way.

Can I change my plan at any time?

Yes. Upgrade or downgrade any time: choose Basic, Pro or Max, sign in, and confirm through secure checkout, then manage or cancel your subscription from Billing in the console, or email support@memroot.dev to switch between paid plans.

Continue the work

Let your next agent start where the last one stopped.

Create free account