MMRY AI Help

Create your account

Every assistant you connect uses the same MMRY AI account, so you create it once. Start your free trial at Get Started. Once you are signed in, that page shows the install steps for Claude Code, OpenAI Codex and connected assistants such as ChatGPT. The same steps are below, by assistant.

When your account is created, MMRY AI emails you a welcome message with a link to connect your first tool.

If your organization already has an account, ask your administrator for an invitation. See Invite teammates.

Install MMRY AI in Claude Code

You need Claude Code installed and an MMRY AI account. On Windows, first install Git for Windows with its default options; the plugin uses the bash shell it provides.

1. Open Claude Code:

   claude

2. Install the plugin with these two commands. The second one answers (no content), which means it worked.

   /plugin marketplace add MMRY-AI/mmry-plugin
   /plugin install mmry@mmry-plugin

3. Type /exit, then reopen Claude Code with claude.
4. Connect the plugin to your account. Claude sees that MMRY AI is installed and starts setup for you. To start it yourself, type:

   /mmry:setup

   Setup opens your browser at mmryai.com. Sign in with your MMRY AI account and choose Authorize this device, and setup saves the connection. If the browser does not open, setup prints a link to paste.
5. Restart Claude Code: type /exit, then reopen it with claude. From the next session, your Foundation directives load as each session starts and Claude searches your other memories when you ask.

MMRY AI is then active. On your first session, Claude helps you create your first memories. Type /mmry:help at any time for a quick reference.

Slash commands in Claude Code

You do not need to memorize these. Talking naturally works too: "remember this" or "what do you remember about X".

Get help in Claude Code

Type /mmry:help for a quick reference. To report a bug or ask for a feature, run /mmry:feedback; your assistant asks a few questions and sends the report to the MMRY AI team.

Set up MMRY AI in Codex

You need the Codex command-line tool, a bash shell and an MMRY AI account. On Windows, install Git for Windows with its default options and run the commands below in the Git Bash window.

1. Install the plugin. In a terminal, run:

   codex plugin marketplace add MMRY-AI/mmry-plugin
   codex plugin add mmry@mmry-plugin

2. Trust the hooks. Restart Codex. The first time it starts after installing, Codex shows a review of MMRY's hooks. Choose Trust all and continue. The hooks are what load and save your memories.
3. Sign in. Run MMRY setup yourself, in a terminal:

   bash "${CODEX_HOME:-$HOME/.codex}/mmry/setup/mmry-setup.sh"

   It opens a browser so you can sign in at mmryai.com. If the browser does not open, the script prints a link to paste.
4. Restart Codex once more. From the next session, your Foundation directives load as each session starts and Codex searches your other memories when you ask.

Use MMRY AI in Codex

• Ask in plain words. Say "remember this", "what do you remember about X" or "make that last one private". Codex turns MMRY's commands into skills your assistant uses for you.
• Approve access once per conversation. The first time your assistant saves or searches in a conversation, Codex asks you to approve network access and access to MMRY's folder. Approve it once and MMRY works for the rest of that conversation. You can allow it permanently in your Codex settings instead, which widens the sandbox for everything the assistant runs.
• Reload when you need to. Ask your assistant to reload your memories, for example after a Foundation directive changes.
• Formations. Ask your assistant to start, join or report to a formation. A formation message reaches Codex the next time anything happens in that session.

Get help in Codex

Ask your assistant to send feedback to MMRY AI. It has a script for it and asks you for the details. You can also write to info@mmryai.com.

Connect MMRY AI to ChatGPT

1. In ChatGPT, open Settings, then Connectors. Turn on Developer mode first if ChatGPT asks.
2. Add a custom connector with this address:

   https://mmryai.com/mcp

3. ChatGPT opens a sign-in page on mmryai.com. Sign in with your MMRY AI account and approve access.

Shared memory for ChatGPT

Connect MMRY AI to Claude on the web or desktop

1. In Claude, open Settings, then Connectors, and add a custom connector.
2. Enter this address:

   https://mmryai.com/mcp

3. Sign in with your MMRY AI account on mmryai.com and approve access.

On a Claude Team workspace, an owner can add the connector once for everyone. Each member then signs in to their own MMRY AI account.

Shared memory for Claude

Connect another assistant

MMRY AI's connector uses the Model Context Protocol (MCP) with sign-in through mmryai.com. The ChatGPT and Claude steps above are the ones we have confirmed. If your assistant lets you add a custom MCP connector by its address, add https://mmryai.com/mcp, sign in, and approve access. Tell us how it goes at info@mmryai.com.

Connector reference for developers

Use MMRY AI in a connected assistant

• Ask in plain words. "Remember this", "what do you remember about X" or "load my MMRY memories".
• Choose who sees each memory. Your assistant can save a memory as private, for a memory group or for your organization, and change that later.
• Foundation directives travel with each MMRY response, so they reach your assistant whenever it uses MMRY in a conversation.
• Formations. Your assistant can list the formations running on your account, join one, tell the group what it is doing, and close one out when you are its lead, set its objective, or administer the account. A formation is started from Claude Code or Codex. Messages reach your assistant the next time it uses any MMRY tool.

Get help from a connected assistant

Write to info@mmryai.com with the assistant you use and what you saw. If you are on a team, your account administrator can often help first.

Save a memory

Tell your assistant something worth remembering, for example:

• "Remember: our deployment branch is develop, production is main."
• "Save this: we use TypeScript with strict mode."
• In Claude Code, run /mmry:save with no argument and it saves the most important takeaway from the recent conversation.

Your words are stored as you wrote them. The next time you start a conversation in any connected tool, that memory is available.

Find something you saved

Ask your assistant: "what do you remember about the billing schema?" It searches your memories and answers from them. In Claude Code you can also run /mmry:search with keywords. In Claude Code and Codex, your Foundation directives load at the start of every session, and your assistant finds everything else you saved by searching.

See which agent saved a memory

When a named agent saves a memory, MMRY AI keeps the agent's name with it, so a team running several named agents can see which one contributed what. The name is kept beside your words, not written into them.

• In Claude Code, the name is picked up automatically when a session is started as a named agent with claude --agent <name>.
• In Codex, and anywhere else the plugin runs, set MMRY_AGENT_NAME in the environment your assistant runs in.
• In ChatGPT, Claude and other connected assistants, your assistant can include its name when it saves, when it runs as a named agent or you have given it one.

The name comes back with the memory when a connected assistant loads or searches your memories, and every memory in your export carries it. A name can be up to 100 characters.

Correct or forget a memory

To correct a memory in Claude Code or Codex, tell your assistant what is right. It finds the memory that is wrong and saves the correction in its place: the earlier memory stops appearing when your assistant loads or searches your memories, and MMRY AI keeps it on record. The correction keeps the earlier memory's visibility, so a private memory stays private and a memory shared with a group stays with that group. If your assistant cannot tell which memory is wrong, it saves the correction and tells you the earlier memory is still there, and you can ask it to forget that one.

Your assistant tells you how each correction went. Usually the correction replaces the earlier memory. Sometimes the correction is saved but MMRY AI cannot confirm that the earlier memory was retired: your assistant tells you the earlier memory may still be active, and you can ask it to forget that one. When a correction is refused, nothing is saved, the earlier memory stays as it was, and your assistant tells you why: the earlier memory was not found, is not active, or is not one you can see; it is a Foundation memory; the correction asked for a different visibility from the earlier memory's; or MMRY AI could not save the correction. A correction cannot replace a Foundation memory. Your Foundation directives are changed by an account administrator in the portal.

In ChatGPT, Claude and other connected assistants, tell your assistant what is right and ask it to save the correction. It can make the earlier memory private if you saved it, and you can remove it in Claude Code or Codex.

To remove a memory, ask your assistant in Claude Code or Codex to forget it. Your assistant deactivates the memory: it stops appearing when your assistant loads or searches your memories and in exports, and MMRY AI keeps it on record.

In ChatGPT, Claude and other connected assistants, your assistant can make a memory you saved private, so only you see it. Removing a memory is done in Claude Code or Codex.

Your Foundation directives

Foundation directives are your standing instructions: the rules you want your assistant to follow in every conversation. In the portal they are listed as Foundation Memories, and an account administrator keeps them there. How they reach your assistant depends on where you use MMRY:

• In Claude Code and Codex, they load at the start of every session and the whole set is sent to your assistant again with every prompt.
• In ChatGPT and other connected assistants, they travel with each MMRY response. They reach your assistant whenever it uses MMRY in a conversation.

In Claude Code and Codex:

• They arrive in full. A set of up to about 57,000 characters reaches your assistant whole, with every prompt. Codex counts bytes rather than characters, so there a set with many accented or non-Latin characters fits a little less.
• A larger set is delivered as a complete copy. Your assistant is pointed to a complete copy and asked to read it before answering. You are told when that happens, and your assistant may need your permission to read the file.
• They are prepared once and read back quickly. Your set is checked and prepared when it changes, and each prompt after that reads the prepared copy, so your directives arrive in time even on a busy machine.
• Your assistant always knows when a turn ran without them. If a very busy machine cannot apply them in time on a prompt, your assistant is told that turn ran without them, so it does not claim to be following them, and the next prompt applies them again. If the stored copy is damaged, that turn runs without them and you are told so in plain words, with what to do.
• You can check. Run /mmry:foundation-status (in Codex, ask your assistant) to see whether they reached your assistant on your most recent prompt.

How big your Foundation set should be

We suggest keeping the whole set under about 10,000 characters, around 2,500 tokens. That is a recommendation, not a limit: nothing stops you going further and nothing you write is trimmed.

An account administrator can see the size of the set in the portal: sign in at your account and choose Foundation Memories. Above the list, the page shows the total in characters and tokens, how many memories that is, and when the set is above the recommendation. Roughly every four characters is one token, and the figure updates as you add, edit, deactivate or restore a memory.

What a larger set costs. Each time the set is sent, it uses some of your assistant's working space, the room it has to hold your conversation, and it costs tokens from your AI assistant's own usage, the plan or account you use it with, not from MMRY. In Claude Code and Codex that happens on every prompt. In connected assistants such as ChatGPT it happens whenever your assistant uses MMRY. A smaller set leaves more room for your conversation.

10,000 and 57,000 are different numbers. About 10,000 characters is the size we suggest. About 57,000 is how much Claude Code can deliver whole on a prompt; above that, your assistant is pointed to the full copy and asked to read it before answering, and the page says so when your set is that large.

Invite teammates

An administrator invites people from your account, in the Invitations panel. Each invitee gets an email link to set a password and join your organization, and a welcome email once they join. Sending and revoking invitations does not use credits.

You can add as many people as your plan has seats. Active teammates and outstanding invitations both count. When every seat is in use, MMRY AI shows how many are in use and asks you to add seats before inviting anyone else.

To add or remove seats, an administrator chooses Change next to Seats in the Subscription panel on your account page, enters the new number, and confirms after seeing the price change. See Subscription and billing.

Who can see your memories

You can set visibility when you save, or ask your assistant to make a memory private. To change your own default for every new memory, open your account and use the Memory Sharing panel: pick Organization, Private, or one memory group, then Save. It is the same setting as the /mmry:visibility command in Claude Code, so changing it in either place changes both.

Share with part of your team: memory groups

A memory group, such as "Backend" or "Leadership", shares memories with only its members. An account administrator creates memory groups and chooses their members from your account, under Quick Actions, then Memory Groups. Save a memory with Group visibility for that group, and only its members see it. If you are not an administrator, ask one to create a group or add you to it.

Start or join a formation

In Claude Code, start a formation in the first session and share the id it prints:

/mmry:formation start "migrate the billing schema"

In each other session, join by that id:

/mmry:formation join 42

In Codex, ask in plain words: "start a formation for the checkout work" or "join formation 42". A connected assistant such as ChatGPT can list the formations on your account and join one; a formation is started from Claude Code or Codex.

A formation is private to your organization. People who share a memory group with its members can join, and your own other windows always can. Each window joins for itself.

Formation commands in Claude Code

Reporting to the lead

Every member of a formation is given two rules when it joins, and every assignment it receives repeats them. They read the same in Claude Code, Codex and a connected assistant:

Report to the lead at every stopping point, with a message directed to the lead: when you finish, when you are blocked, when you are waiting, and when you stop with work still running, saying what is still running. Never hold finished work back for a slow check; report it and say the check is still running.

When you receive an assignment, a brief or a question, acknowledge it before you start, with substance: what you understood, anything you cannot do, and when to expect the result. If the job is short, fold that into your first update. Findings and status updates need no acknowledgement.

So the person running the job does not have to watch every window or repeat this in each brief. An acknowledgement is more than a read receipt: it is where a member says what it understood, so a misunderstanding is caught before the work starts.

When formation messages arrive

A message waits until it is read and is delivered once. A Claude Code session picks messages up between the steps of its work, while it sits idle, when its person types, and when it next starts. Keep Claude Code up to date so a message can reach a session that is sitting idle. A Codex session picks them up the next time anything happens in it, and a connected assistant picks them up the next time it uses any MMRY tool.

In Claude Code and Codex, formation messages arrive under the heading "MMRY AI Formation Transmission", so they stand apart from the rest of the conversation. When a message wakes a Claude Code session that was sitting idle, the session shows "MMRY AI: Formation Transmission".

When something has to be acted on now, ask that session to run /mmry:formation status (in Codex or a connected assistant, ask it to check the formation). That delivers everything outstanding straight away.

Staying ready for messages while in a formation

In Claude Code, a session in a formation keeps listening for messages for as long as it stays in the formation, however long it sits idle. About every half hour of quiet, MMRY renews that watch: the session wakes briefly, replies "Still listening.", and goes back to waiting. That turn is expected and needs nothing from you.

While your session sits quiet in a formation, it checks in about every half hour so it picks up new messages within about a minute of their arrival, without you typing anything. Each check-in is a short turn with a two-word reply. Claude reuses the conversation it already has loaded, so almost all of each check-in is re-reading rather than new work, and it counts toward your Claude Code usage like a normal turn. To stop the check-ins, leave the formation with /mmry:formation leave.

The watch ends when the session leaves the formation, joins a different one, or the formation is closed out. A session that is closed, or on a machine that has gone to sleep, picks its messages up when it next runs. Nothing is lost while it waits.

This describes the MMRY AI plugin for Claude Code from version 2.10.2 onward. A plugin installed from the marketplace updates itself when Claude Code restarts; see Keep the plugin updated automatically.

Sessions that are not in a formation skip the formation check entirely, so being outside a formation adds no wait to your work.

Check whether a member has read your message

When you send a message to one member, you can check whether it has reached them:

• In Claude Code, run /mmry:formation roster. Under the members it lists each message this session sent to one member, newest first, marked READ with the time or NOT READ YET. The read time is shown in UTC.
• In Codex, ask your assistant to show the formation roster.
• In ChatGPT, Claude or another connected assistant, ask it to check the formation. Its status lists the messages it sent to one member and whether each has been read.

Read means the message has been shown to that member's session. It does not mean they have acted on it yet. Read status belongs to the session that sent the message, so check it from that session: another session, even one of yours in the same formation, does not list it. If the member a message went to has since left the formation, the roster keeps its READ or NOT READ YET mark and adds (has since left). A message sent to the whole formation has no read state, because it has no single reader. If a message shows as not read yet and it matters now, ask that member to run /mmry:formation status, or to ask their assistant to check the formation.

Close out or leave a formation

When the job is done, the lead or an administrator reads /mmry:formation report and closes the formation out with /mmry:formation debrief "summary". Close-out always saves the formation's record and the summary as memories the team can read later. If it cannot finish, it says why and whether trying again helps.

To step away yourself, run /mmry:formation leave. It affects only that session, and the formation carries on for everyone else.

Your account dashboard

Manage your account at your account. From there you can see your users, manage memory groups, review your subscription, and export your memories.

Your records

The record types you keep, and the records in them, are at your records. A record type is a structure you have asked your assistants to keep for you, such as a migraine log or a set of tracked tasks, and a record is one entry in it.

You can correct a record there yourself: open the type, find the record, and fix the value. The change is true for every assistant on your account the moment you save it. Every change is kept, so you can also see what a record held before. A record's history is exactly as private as the record. To design a new record type or retire a record, ask an assistant.

In Claude Code, ask your assistant to keep track of something, such as expenses, migraines or job applications. It sets up a record type for it, files what you tell it as records, and answers questions about the values: how many, which ones, since when, sorted by what. It can also correct a record and show you every change a record has been through, oldest first.

Free trial

New accounts start with a 14-day free trial, with the same credit pool as the paid plan and nothing charged. When the 14 days end, your trial converts to the paid plan, and that day becomes your monthly renewal date. If your team uses up the trial's credit pool before then, your paid plan starts that day instead. Every account administrator gets an email when the paid plan starts. An account administrator can cancel before the trial converts from Manage Billing on your account, and nothing is charged. See the Terms of Service.

Subscription and billing

Billing is handled by Stripe. An account administrator opens Manage Billing on your account to update payment details, see invoices, or change the plan in the Stripe customer portal.

Add or remove seats. An account administrator changes the number of seats in the Subscription panel on your account: choose Change next to Seats, enter the new number, and the page shows the price change before you confirm. You are charged the price you previewed; if it changes before you confirm, the page asks you to preview the change again. The new count takes effect at once. Adding seats charges the prorated amount for the rest of the current billing period; removing seats credits the unused portion to your next invoice. You cannot go below the number of seats in use, which counts active teammates and outstanding invitations, so remove people or revoke invitations first. During a free trial, seats change at once and the change itself is not charged; your paid plan starts when the trial ends or when your team uses up its credit pool, whichever comes first. You can choose up to 1,000 seats. If seats cannot be changed, the reason and the next step are shown next to Seats instead. For example, while a payment is past due, update your payment method from Manage Billing, then change seats.

Credits and renewal. Your plan is billed per user per month, and every user adds credits to your team's shared pool; the current price and the credits each user adds are on the pricing page. Every request your AI tools make to MMRY AI on your account uses one credit from the pool, whether it saves, loads or searches memories, exports them, or sends a formation message; signing in, inviting teammates, and managing billing, users and groups do not use credits. The pool refills at every monthly renewal. If your team uses the whole pool before the month ends, your plan renews early: every user is billed for the next month, the pool refills, and that day becomes your new renewal date. On a free trial, using up the pool starts your paid plan the same way. If the renewal payment does not go through, requests are refused until an account administrator updates the card in Manage Billing. See "Your credits are used up" messages.

Renewal emails. Every account administrator gets an email when the plan renews early and when a trial converts to the paid plan. It states the amount billed, the credits in the pool and when the plan renews next.

Cancel your subscription

An account administrator can cancel at any time from Manage Billing on your account. Your subscription stays active until the date shown on your account page: the end of the current billing period on a paid plan, or the end of the free trial on a trial. Until that date, you keep saving, loading and searching memories. On a paid plan, export your memories as JSON before then if you want a copy. Downloading all of your memories is available on a paid plan.

Your privacy choices

Every page of the site has a Privacy choices link in its footer. It opens the Your privacy choices section of the privacy policy, where you can turn analytics and advertising measurement off, and back on, for the browser you are using. Your choice is stored in that browser only, so repeat it in each browser you use. It applies on every page, including the pages you use after signing in.

Browser signal. If your browser sends the Global Privacy Control signal, MMRY AI treats it as an opt-out, so nothing is sent to any measurement tool from that browser and there is nothing more to do.

Consent notice. If you appear to be visiting from the European Economic Area or the United Kingdom, a notice asks you before any measurement tool loads. Choose Allow measurement or Decline, and change your answer at any time from the Privacy choices link.

Export your memories

Open your account and choose Export Memories under Quick Actions. Your memories download as a JSON file.

Downloading all of your memories is available on a paid plan. During a free trial, your account page offers an upgrade link in its place.

Delete your data

To remove one memory, ask your assistant in Claude Code or Codex to forget it. MMRY AI deactivates the memory: it stops loading and stops appearing in searches and exports, and MMRY AI keeps it on record. In ChatGPT, Claude and other connected assistants, your assistant can make a memory you saved private, so only you see it.

To have your account and all of its data deleted, email privacy@mmryai.com.