Memory

Your bots remember things about you so you don't repeat yourself: facts ("my dentist is Dr. Lee"), preferences ("when giving news, use 3 bullets with links") and open tasks ("send Sam the deck by Friday"). Memory belongs to you and is shared by all your bots. You can see, correct and remove every bit of it in the Memory tab.

• Only one bot. A memory can be kept by one bot alone: ask it ("just for you: …") and it saves it with only_me, or switch a bot-learned memory between all bots and only <bot> on its card. Your other bots don't see it in their standing memory or searches; a new version stays with that bot.
• The household. Lists and plans you share with people you choose (a shopping list) live in the Household: see Bots and chats. Bots in household chats get it as its own block next to standing memory, and memory_search searches it too (marked household). What a guest adds arrives as a suggestion in your Needs you.

How bots use it #

• Standing memory: your preferences and the most recent facts and tasks go into every reply's instructions, within a size budget (about 2,000 tokens). Pin a memory to keep it there however much else there is. The Memory tab shows how full it is ("Standing memory 67% full · 12 memories in every reply").

• Search: everything else is found with memory_search when a question needs it ("who's my dentist?").

• Saving: a bot saves with memory_store when you tell it something worth keeping. To correct a memory it saves a new version; the old one stays under earlier version.

• End dates: something true only for a while ("I'm in Lisbon until Friday") is saved with an end date. Through that day it's in standing memory; after it, it drops out, and a search still finds it marked ended. A new version keeps the end date. Change or remove it on the memory (until … · no end); the tidy-up may suggest forgetting ended ones.

• Earlier conversations: what you said but nobody saved ("which hotel did I book in Lisbon?") is found with chat_search, over all your past conversations with your bots, optionally between two dates ("what did I say about the hotel last week?"). Your own words count as yours; a bot reply that was written after reading a web page is marked as such, and finding one counts as reading outside content again. Incognito chats are never stored, so never found.

Three kinds: preference ("convention": a standing instruction), fact ("knowledge") and task (an open commitment).

You stay in charge #

• Preferences need your own words. A bot saves a "When …, …" preference directly only when your message in that reply states the rule ("from now on…", "always…", "I prefer…") and it matches what's saved. A habit a bot works out on its own becomes a suggestion you accept or dismiss.
• No secrets. Memory refuses passwords and API keys; use saved logins for those (Tools and safety).
• Where it came from. Mimir records who wrote each memory (you; which bot, in which run; a tidy-up; an import) and never trusts the model's own claim. A memory a bot saved after reading web pages or tool results is marked after reading outside content, and your bots treat it as unconfirmed, so a planted instruction doesn't read like something you said: it stays out of their standing memory (a search can still find it, labelled, and then that reply counts as having read outside content). The mark sticks: a new version of an unconfirmed memory, a bot's proposal made after reading a page, or a tidy-up suggestion drawn from such a conversation is flagged on its card, stays unconfirmed when you accept it (also through "Accept all"). Looks right clears the mark. The exception is a suggestion that says what you wrote yourself ("No, use metric units only" becoming a metric-units preference): a page can't put words in your messages, so it stays yours even if a bot searched the web in that conversation. Memories waiting for Looks right are listed under Needs you, so they don't sit unseen (and unused).
• All at once. With two or more suggestions, Accept all and Dismiss all answer them together; with two or more memories waiting for Looks right, All look right confirms them after you've read them.

Files #

Next to memories, each chat has a small workspace of text files your bots keep there with file_write: a packing list, running notes, a draft, research too long for a memory. They're under Memory → Files (open, or delete); the folder is files/chat-<id>/ in Mimir's data folder. Rewriting a file keeps the previous version, deleting one asks, and a file written after reading a web page stays marked as outside content when a bot reads it again. Like memory, it belongs to the chat, so a group chat can't read your private chat's files.

Changing and removing #

• Edit a memory in place; the previous wording is kept under earlier version.
• Under a reply: a quiet line says what it used ("2 memories · 1 skill · read 3 pages"). Open it for the memories recalled for that reply, each with Edit and Forget, plus the skills it read and the pages it opened. In the messaging apps, /why lists the same for the bot's last reply. Standing memory is in every reply, so it isn't listed.
• Move to background takes it out of standing memory without deleting it (search still finds it).
• Forget hides it but keeps it under History; Erase (in History) deletes it and its earlier versions for good.
• Import takes text pasted from another assistant ("List everything you remember about me") and turns it into suggestions, skipping what you already have and anything that looks like a key or password. When Claude Code, Hermes or OpenClaw is set up where Mimir runs, Import also offers From Claude Code (etc.): one click reads its memory files (~/.claude/CLAUDE.md and Claude Code's per-project memory; Hermes' USER.md and MEMORY.md; OpenClaw's workspace USER.md, MEMORY.md and memory/) into suggestions. That agent wrote them, often after reading repos, pages or mail, so they count as outside content: new memories only (never a change to one you have), unconfirmed even after you accept them, until you mark one Looks right. Its own SKILL.md files go into skill proposals (Hermes' bundled skills are left out), and its cron jobs into routines that start paused: resume one to accept it. Imported routines get no extra tools and count as outside content, like a routine a bot made after reading a web page. Only cron schedules come over (not one-offs or intervals), on Mimir's clock.

Suggestions from your email and calendar #

Settings → Get to know me shows up once your email or calendar link is connected, and runs once. Your first bot's account gets the senders, subjects and dates of the last 30 days of your inbox and your calendar from a month back to two weeks ahead (titles and who's invited), with no tools: it can't read an email's text, write, send or mark anything. It proposes up to 15 new memories (people you work with, recurring meetings, upcoming deadlines). They arrive as suggestions in the Memory tab, never as memories. Because they come from other people's words, they only add (never change or retire a memory you have) and stay unconfirmed even after you accept them, out of the bots' standing instructions until you mark one Looks right.

The daily tidy-up #

Once a day, after a quiet half hour, your first bot (setting memory_bot) reviews memory against recent conversations and suggests changes: remember a new fact or habit, update an outdated one, forget a finished task or a duplicate, move something rarely relevant to the background. Suggestions appear at the top of the Memory tab with a reason; nothing changes until you accept. Tidy up now runs it on demand; a dismissed suggestion isn't proposed again for 90 days, and neither are lookalikes: one like a suggestion you dismissed says so on its card ("You dismissed a similar one on 3 Oct"), and after two similar dismissals that kind isn't shown at all (imports excepted). It runs on your first bot's AI account; Settings → General → Model for the memory tidy-up can point it at a cheaper model (e.g. haiku, gpt-5-nano), which matters on pay-per-token keys. The setting belongs to the kind of account the memory bot uses (a Claude model name means nothing to a local model); a failing tidy-up retries after an hour.

It learns from corrections. When you push back on a reply ("no, I said one sentence"), stop a reply, or deny an action, that's noted. So is asking nearly the same thing again within ten minutes (in any language), a hint the reply missed, though weaker: it never counts against a memory, and the tidy-up uses it only next to a correction or your own words. Correcting a routine's result ("skip the crypto stuff") is noted with the routine it's about, so the suggestion is a preference for that routine ("When doing the tech news briefing, …"), not for everything. Two of those bring the tidy-up forward (after 10 quiet minutes, at most hourly). It proposes a "When …, …" preference from them only when the same correction happened twice or you stated the rule yourself, and it quotes the evidence. Each reply remembers which memories it looked up; when you correct it, those memories count a corrected after use, and so do standing preferences your correction is about (they shape every reply, so they're never "looked up"). A memory that keeps steering replies wrong shows up with a proposed fix. Memories reach your bots with the date they were saved, so an old fact can be weighed against a newer one.

A correction carries over to similar requests. When you ask something like a request you corrected before (most of the same words), the bot sees what you asked then and what you said ("Too long. Three short bullet points, no intro."), so it doesn't make the same mistake again, even before any preference is saved.

New facts update old ones. The tidy-up checks each memory saved since its last run against the rest. When a new one makes an older one outdated, directly ("moved to Utrecht" vs "lives in Amsterdam") or as a knock-on ("broke my leg" vs "cycles to work"), it suggests updating the older one or moving it to the background, naming the new memory. It never suggests forgetting one for this, so the history stays.

Check before accepting. On a suggestion, Check replays your past corrections that look related, each once with the change and once without, and a judge that isn't told which is which says whether each reply still has the problem you corrected. It says, for example, "Fixes 2 of 3 past corrections about this; 1 was already fine without it", so a change your AI no longer needs shows as such. It uses your AI account for a few short replies (sent all at once, then judged), so it runs only when you press it.

Accepted preferences have Check too (on the memory). It replays on your memory bot's own model, not the cheaper tidy-up model, so after switching models you can see whether a preference is still needed: one that fixes nothing the new model doesn't already get right offers Move to background.

Learned this week #

The Memory tab lists what was saved without you typing it (by a bot, an accepted tidy-up, an import) since you last looked, each with Undo (which also brings back the version it replaced), plus a trend line: corrections per 10 messages against the weeks before, corrections that repeated an earlier one (the same kind of request corrected again: what learning should bring down), "I told you" corrections (something a bot should have remembered), denied and stopped actions, failed actions, suggestions accepted, and how many learned memories you took back with Undo (how often learning gets it wrong). Once a week at your digest time, each Telegram, Slack or Discord chat gets the same list (a new chat's first one comes a week after it starts); reply /undo 2 to take one back.

A memory a bot saved shows the message of yours it came from ("You said: 'I might start going to the gym on Mondays'"), so a passing remark saved as a firm fact ("goes to the gym every Monday") is easy to spot.

Skills #

Procedures your bots follow step by step (the ones they learn from tasks that went well, the ones you teach, and the ones from Mimir's library) have their own page: Skills.