Conversation Search Across Every Octo: Octomind Cloud Update

Say an octo, one of the AI assistants you run on Octomind Cloud, worked out what a supplier charges three weeks ago. You remember the answer exists. You don't remember which octo found it, which conversation it was in, or what word it used. Until today, you scrolled.

Conversation search is now in Octomind Cloud. Press Cmd+K (Ctrl+K on Windows and Linux), type what you remember, and it searches your whole chat history across all your octos. It forgives typos and finds messages by meaning when your words don't match. This post covers everything new in Octomind Cloud since our October 31 update.

Conversation search from one menu

Cmd+K and the search icon in the sidebar open the same command menu. Before you type, it lists actions, your recent conversations and your octos. As you type, it narrows those and adds history: matching messages, each with its octo, its conversation and a snippet with your words marked.

Specialists are the agents an octo hands a job to, and events are scheduled and webhook messages.

Press Enter on a result and the conversation opens at that message, with your words marked. If the result is a specialist's answer, that answer opens unfolded. Inside a specialist's log, the find in log box searches that log and counts the matches.

The menu also starts work. Type ask, an octo's name, a colon and your message, such as ask Research: compare these three suppliers, then press Enter. It starts a new conversation with that octo and sends the message.

What chat search covers, and what it skips

Past conversations are already searchable, and new messages become searchable within seconds.

• Covered: your messages and the names of files you attached, your octos' replies, what an octo asked a specialist and what came back, and scheduled and webhook messages.
• Skipped: tool output and anything else an octo did along the way. A file it read or a command's output won't show up unless the octo wrote it in a reply.
• Long messages: searched up to their first 256 KB.
• Secrets: text that looks like an API key, token or password is replaced with [redacted] before a message is indexed, so results don't show it. The conversation itself keeps the original.
• Your account only: every query is filtered to your account on the server.

Searching chat history is part of Pro, Max and Team, with no extra charge. On Free, the menu still works for actions and getting around the panel, without history.

Keyword search vs semantic search: how hybrid search finds what you meant

Keyword search alone fails the moment you remember the idea and not the word. So each search runs three ways:

1. Exact words, with the last word completed while you're still typing.
2. Close spellings (fuzzy search) if nothing matches. relase finds "releases", and the menu says it's showing close spellings.
3. Meaning (semantic search). A small model, all-MiniLM-L6-v2, turns each passage (a message, or a piece of a long one) into numbers that capture what it's about, so a query can find a passage worded differently. The model runs on CPU inside Manticore Search, the search engine behind this, so nothing leaves our servers for it. Results found only by meaning carry a related label.

Keyword and meaning results merge into one ranking, which is what hybrid search means. Messages that contain your words always rank above related ones, and among them, a message that also matches in meaning ranks higher. Any of your words that appear are marked, even in a related result.

We tested it on 80 labeled queries against 296 indexed passages. That's a small set, so read these as a sanity check:

Of 15 queries about things that weren't there, 1 returned anything. On that test index a search took 14 ms at the median and 23 ms at the 95th percentile on the server. In production, a three-word query against the 13,700 indexed passages in my own account took 21 to 28 ms in the search engine.

See what each octo spent

The Usage page now breaks spending down by octo, and an octo's total now includes media, web search, the cloud browser and voice, not only models. Expand an octo to see each: model usage, media, web search, browser starts, browser time, voice in and voice out. Choose 24 hours, 7 days or 30 days, or click a bar to narrow to one day or hour. Non-model spending counts from this release on, so older days show models only. Machine time is still billed per machine, not per octo.

A conversation's spending view now has a spending by day chart, in UTC days, for the last 7 days, the last 30 days or all time.

Try Pro for a week for $12

If you're new to paid plans, Billing now offers try pro for 7 days: $12, with $12 of usage included. It's for personal accounts that haven't subscribed before. After day 7 it renews at $50 a month, unless you cancel under manage subscription first.

Don't want to wait out the week? Click start pro now: it charges the $50 today and starts your month, and any unused trial usage moves to your credits. Plans are on the pricing page.

New specialists and an AI visibility report

Eight new specialists can take jobs from your octos: literature review, thesis, form filler, store copywriter, store policy, Discord builder, community moderator and vehicle inspector. New skills, instructions an octo loads for one kind of task, cover budget meals, gift shopping, product reviews, budget travel, packing, road trips and vehicle maintenance. Another skill audits a store's product feeds and pages for AI shopping assistants and lists the fixes.

A new workflow, AI Visibility Report, asks the AI assistants you name, such as ChatGPT, Claude, Gemini or Perplexity, a fixed set of buyer questions, records which products they name and which sources they cite, compares that with the last run, and drafts an action list for you to review. It drafts only. Nothing gets posted or sent.

New and retired models

• Retiring: Claude Fable 5.1 on November 17. Claude Fable 5.2 replaces it at $8 / $40 per 1M tokens, down from $10 / $50, with cache reads at $0.20 instead of $0.25 and the same 1M window.
• Retired: DeepSeek V4 Flash on November 1, Mercury 2.5 on November 4, and GPT 5.6 Terra today. GPT-6.1 Terra replaces it at $1 / $5 per 1M tokens, down from $2 / $12.
• Corrected: Inkling and Inkling Small have a 256K context window. We listed 1M.
• Evaluation models: eleven more of the small models that answer yes/no questions, from $0.01 to $0.24 per 1M input tokens.

Each model has its own page on the models page.

Smaller things

• Quotes link back. Click a quote in your message and the chat scrolls to the line you quoted and highlights it.
• Slow models get longer. A model call can now run for 10 minutes before it counts as failed, up from 5, so models that think for a long time get the time they need.
• Conversations compress later. When a conversation grows long, Octomind summarizes its older part. That now starts at 90,000 tokens, up from 70,000, so more of a conversation stays word for word, unless you changed the setting yourself.
• Cloud browser closes sooner. It now shuts after 2 minutes idle, down from 5. You pay for it by the minute, so idle time costs less. When a site needs your sign-in, the octo shows a button that opens the page and an I've finished signing in button to press when you're done.
• No more false alarm on a sleeping octo. Your first message to an octo that was asleep could end in "no response from the agent" even though the octo had answered. Fixed.
• Windows install uses Scoop. Run scoop bucket add muvon https://github.com/Muvon/scoop-bucket, then scoop install muvon/octomind.
• New machines run Octomind 0.70.1. The 0.70.1 post covers provider setup in one screen, and the 0.68.0 post covers supervisor advice.

An iPhone app is on the way

We're building a native iPhone app for your octos, and conversation search is in it. It isn't released yet. When it is, you'll see it on the cloud page and hear it from us here.

Search for something you gave up on finding, and tell us what it missed. Post in Discord or reply to our newsletter. The query and what you expected to see are usually enough for us to fix it.

— DK