Skip to content

Mimir v0.6.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 23 Oct 19:57
· 50 commits to master since this release
v0.6.0
7de8c76
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Email

  • Bots say they're an assistant: an email a bot sends to someone you haven't written to yourself ends with "Sent
    by my assistant" (change the line, or turn it off, in Settings → Email). On for new installs. A reply to a
    bot's email that's personal, off-topic, or asks you to commit to something (a meeting, money, a contract) isn't
    answered: it comes to you in Needs you with a one-line summary.

Memory

  • Better recall: an old memory that matches your message exactly is no longer pushed out by a newer one that
    only shares a word, preferences don't fade with age, a match in a memory's title counts for more than one in its
    text, and long messages are searched on their key words rather than their first sixteen. Unconfirmed memories no
    longer take recall's six places, and ended or background memories rank lower in it.
  • Sensitive topics stay out of memory: the tidy-up, imports and Get to know me no longer suggest memories about
    your health, religion or beliefs, politics, sexuality or finances (account numbers too), unless you say "remember
    that…" yourself. Turn on Settings → Memory → Remember sensitive topics to let them in. See
    Memory.
  • Import from Codex: Memory → Import → From Codex brings in ~/.codex/AGENTS.md as memory suggestions and
    your own Codex skills as proposals, under the same rules as Claude Code, Hermes and OpenClaw: outside content,
    nothing active until you accept it.
  • Memory as files: Settings → Memory can keep your memories as Markdown files too, in a folder you pick: one
    file per memory, a folder per chat, with frontmatter (type, chat, bot, dates). Open it in Obsidian, grep it, keep it
    in git. What you edit, add or delete there waits in Needs you as a suggestion; nothing changes until you accept it.
    Guests' and incognito chats stay out, and files already in the folder are left alone.

Talking to your bots

  • Word, PDF and Excel files from bots: ask for "the expense report as a spreadsheet" or "this as a PDF" and the bot
    sends a real file, on every channel. Headings, lists and tables come through; a spreadsheet gets a bold, frozen
    header row. Files hold text only (no links, macros or formulas), and PDFs print only Latin script for now.
  • An allowance per person: Settings → Limits → Each person caps what replies to one guest or family member may
    use in a day, in their own chat and a family group together. Over it, they get a short note instead of an answer,
    and Needs you offers Allow more today. Your own chats are never capped.
  • iMessage, in the Mac app: a bot can answer in one Messages conversation you choose, a note to yourself or a
    chat with a second Apple ID, and never reads any other. Text, photos and voice messages in; replies as plain text,
    with approval cards and questions as numbered lists, as on Signal. Needs Full Disk Access and permission to control
    Messages, and works only while the Mac is awake. Setup is under Connect.
  • Mimir on your phone's home screen: the server's web app can be added to a phone's home screen and opens full
    screen with its own icon. See Self-hosting.
  • Telegram groups: a bot in a group (not a family group) answers only when you @-mention it, reply to one of its
    messages or send a /command. Before, with privacy mode off, it answered every message.
  • Six more skills in the skill library: Deep research (a thorough report from many sources, with a plan you
    watch and every claim labelled), Chart check (trend, levels and indicators for a stock, currency or coin; a read,
    never a trade), Make it sound human, SEO check (keywords, titles, tags, and claims ads law bans, also in
    Chinese), Prompt polish (a rough idea into a clear prompt for a routine, a bot or another AI tool) and Landing
    page     (a modern one-page site from a business's current website). See docs/skills.md.

Moving in

  • Moving from OpenClaw or Hermes: Memory → Import → From OpenClaw now finds a workspace set in
    openclaw.json or OPENCLAW_WORKSPACE_DIR, and brings in every profile (~/.openclaw-<name>,
    ~/.hermes/profiles/<name>) too. Its SOUL.md is offered as the bot's role (one click) instead of becoming
    memories, imported memories say which agent they came from, and a cron job set for another time zone says so on
    its paused routine. See Getting started.

Routines

  • A second look: a ready-made quiet check in the Routines tab, off until you add it. On each email that looks like
    a confirmation, quote, invoice or booking, it compares the dates, amounts, names and quantities with the earlier
    emails, your calendar and memory, and speaks up only when they differ ("Hotel confirmation says 12–14 Nov; your
    calendar says 13–15"). It only reads; it never replies. Mail rules take a|b for either word. See
    Quiet checks.
  • Claude on your monthly API credits: Max and Team plans include monthly API credits. Save an Anthropic API key
    from your linked Console organisation (Settings → AI keys), and a Claude bot can go on with it when the plan says
    no, in the same conversation, or run its routines on it from the start. The credits stop when they run out: nothing
    is billed. See docs/ai-accounts.md.
  • Started by a Slack message: slack:#deploys from:@sam failed starts a routine when a message in a channel the
    bot is in matches, from a person or another app (a CI bot). Mimir does the matching, so other messages cost nothing.
    "Tell me when someone says it's deployed in #releases" works in chat. See
    docs/routines.md.
  • Started by GitHub: On GitHub in the Routines tab (on a server) gives a routine a webhook for GitHub with an
    event filter, like issues.opened repo:acme/app. Other events get no run, and the run gets a short summary of the
    event instead of GitHub's raw payload.

Tools

  • More tool servers to browse: Excalidraw (diagrams drawn as a card in the chat), Gamma (presentations, with an
    API key) and a new Analytics group with Mixpanel, PostHog and Amplitude, under Settings → Tool servers → Browse.
  • Box, HubSpot, Calendly and Cal.com: in Browse too, with a sign-in that works. Calendly and Cal.com are one
    click; Box and HubSpot don't let apps register themselves, so their entries say how to make one and ask once for its
    client id and secret.

Safety

  • Chat API tokens can't read your email: an app using a chat API token now talks to the bot
    in a chat of the token's own ("Chat API: Home Assistant"), not your built-in chat. It isn't private, so no email,
    calendar, meetings or computer there, and a rule said through it is a suggestion for you to accept. Whoever holds the
    token, or talks to the speaker that uses it, isn't necessarily you.
  • Quiet checks stay off your network: a watched page that redirects to a name pointing at this computer or the
    local network (a home router, a cloud server's metadata) is no longer fetched; before, only addresses written as
    numbers were refused. Quiet checks and web_fetch also read at most 2 MB of a page.
  • Signal: photos, voice notes and files people send the bot are deleted from signal-cli once Mimir has them, and
    ones it ignores (a stranger's, a group's, one too big) are deleted too; before, they stayed in signal-cli's storage
    for good. A download stops at 20 MB whatever size the sender claims, and a user:password@ in signal-cli-rest-api's
    address no longer shows in errors or the bot's status. Connect now warns that web pages
    on the same computer may reach signal-cli-rest-api.

Fixes

  • Signal and WhatsApp: replying to an approval card with a number now works after Mimir restarts. Before, the
    number was taken as a "no" with the number as the reason.
  • Skill imports: a skill imported from another agent or a bot file with the name of one you already have (perhaps
    one all bots share) comes in as "‹name› (imported)" instead of replacing it when you accept. A taught skill that
    would replace one says so on its card.
  • Grok and Claude bots: on Grok Build 1.0.62 and later, a reply no longer waits on grok's background work for
    longer than a turn may take. Claude Code's fork subagents, on by default since 2.1.302, stay off for bots, so a
    reply is one process and its usage is that turn's.
  • Telegram files: a photo, voice note or file over 20 MB gets a reply saying so, instead of nothing. Your own
    photos and files in a group are taken when they're addressed to the bot, like text. Buttons older than 7 days say
    they've expired. Two files saved in the same second no longer overwrite each other, and a guest-mode reply that's
    cut short says so.
  • Signal: the bot's "I heard" echo or a file upload no longer closes an open question's numbers.
  • "Just for you" on a shared memory: saving a memory every bot already has as "just for you" no longer hides it
    from the other bots.

Installers for macOS (Apple silicon), Windows and Linux. The Mac app is signed and notarized by Apple. The Windows installer isn't code-signed yet: choose More info, then Run anyway the first time. How to use Mimir: https://heymimir.app/docs/

Assets 20
Loading