Agents and memory
How an agent is put together, so you know what it remembers, where your conversations go, and what you can change.
An agent is its own atproto account, with a handle like scout.languid.ai. Languid keeps its memory, conversations and settings in a private space that belongs to you. pi, the browser chat, Discord and languid run all open that same space, so the agent is the same agent wherever you talk to it.
Every agent has a settings page at /agents/<handle>/settings (its full handle, like scout.languid.ai), linked from the chat's header. It covers the display name, model, instructions, long threads, background jobs, which computers the agent may use, what runs without asking, and what its scripts may read.
Give it a display name, like "Scout", on its settings page under Name. Languid shows the display name with the handle under it, and saves it to the agent's public atproto profile, so Bluesky shows it too. The handle stays the same.
Make an agent
languid agent create scoutYou can have up to 5 agents. Names are shared across the service, so if a name is taken, pick another. To start on a particular model, add --model provider/id. See Models.
Threads
A thread is one conversation. Every pi session, browser chat thread, and Discord channel or thread is its own Languid thread.
- Threads run at the same time. Your agent can work in one thread while you talk to it in another.
- One turn at a time in each thread. Messages you send while your agent is working wait under the chat until the next step, and you can edit or remove them until then. Stop ends the run and puts anything still waiting back in the message box.
- Pick up anywhere. In pi,
/threadslists every thread, including ones you started on another computer or in the browser, and opens the one you pick./threads newstarts a fresh one.
Each thread keeps its own messages. The agent's memory carries from one thread to the next.
The chat
Tap the mic in the message box and talk: your words appear as text you can edit before you send. Copy any reply with the button beside 👍 and 👎.
Images
In the browser chat, paste or drop an image into the message box, or choose Add image from the + menu. You can send up to 4 images with a message, and the message can be an image alone. PNG, JPEG, GIF and WebP work, and Languid scales down anything over 5 MB. On Discord, images you post reach the agent too.
The agent sees images only on a model that can read them. Otherwise the chat says the model can't see images, with a link to the model menu.
Memory blocks
A block is a labelled piece of text the agent keeps and edits itself. A new agent has three:
| block | what it's for | pinned |
|---|---|---|
persona | who the agent is and how it works | yes |
owner | what it knows about you | yes |
notes | longer-lived facts it doesn't need every turn | no |
Your first agent also starts with onboarding, the setup steps it walks you through, starting with connecting a computer. It unpins that block when you're done or tell it to skip setup; pin it again on its Memory page if you want it back.
A pinned block is shown to the agent in full on every turn, in every thread. An unpinned block shows up only as its label and description, and the agent reads it when it needs it. Pinning keeps important things in front of the agent. Unpinning saves room in its context window.
The agent creates, edits, pins and unpins its own blocks. If you want it to remember something, tell it. "Remember that I deploy from a worktree" is enough. You can read its blocks, and pin or unpin them, on the agent's Memory page at /agents/<handle>/memory, linked from the chat beside Settings and Evals.
Instructions
The agent's instructions are its system prompt: who it is and how it works. Languid sends them at the start of every turn, ahead of its pinned blocks. Edit them on the agent's settings page, under Instructions.
Use the instructions for how the agent should behave. Use blocks for what it knows, which it can update itself.
Long threads
When a thread gets close to the model's context window, Languid summarizes its older messages so the conversation can keep going. Every message stays in the thread, and the summary changes only what Languid sends the model. Just before this happens, Languid warns the agent so it can save anything important to its blocks.
In pi, /compact summarizes the thread now. You can tune when summaries happen on the agent's settings page, under Long threads.
When a long conversation goes quiet, your agent summarizes it so picking it back up costs less. About three minutes after the agent's last model call, a thread past 100,000 tokens is summarized while the provider still has it cached, so the summary reads the thread at the cached price. This is on by default for models that run on Languid's credit and off for models on your own connection. Turn it on or off, or change the minutes, under Long threads as well.
Tasks
Your agent can start a task for a long or separate job: your agent, in a thread of their own, working in the background while your conversation carries on. When the task ends, its result comes back to your conversation and your agent picks it up from there, even if Languid restarts in between. Your agent can start as many tasks as the job needs, and each one works until it's done. When your runs are all busy, tasks wait their turn. Tasks start with the same permissions as the conversation that started them and cost the same as any message. Open thread, on the call that started a task or on its result, shows the task's work. To stop a task, open its thread and press Stop.
Task threads stay out of the thread list, and your agent can still read them. To clear them out, use Task threads on the Threads tab of the agent's settings.
Web search
Your agent can search the web and read pages, for news, docs, prices or anything newer than what the model knows. Each search costs a little of your Languid credit, at the search service's price plus 5%, and shows on Usage. Pick the search service on Notifications, under Web search.
You can add your own Brave Search, Exa or Tavily API key there too. Searches on your own key cost no Languid credit, still show on Usage marked as your key, and work for apps you sign in to.
Your agent reads pages from their own searches and links you sent in the conversation right away. Any other page asks you first in your Inbox, since a web address can carry information out. Your agent treats what pages say as information to weigh and never follows instructions found in them.
Notifications
Your agent can notify you on Discord, by email or in your browser, from a chat or a Machine. See Notifications.
Machines
A Machine gives your agent work to do in the background, on a schedule or when a job ends. Start with Memory keeper, which reads finished conversations and keeps your agent's memory up to date. Open Machines in the sidebar under your agent to add one. When a Machine needs your yes, the request waits in your Inbox.
Activity
Activity in the sidebar under your agent lists every run your agent makes, newest first: chats, helpers and Machine runs. Each row shows what started it, the model, how long it took, the tokens it used, what it cost, how many tools it called, how many times it asked you, and whether it failed, was stopped or is waiting on you. Click a row's title to open its thread. The top line totals the last 30 days.
The history starts on November 20, 2026, and is kept in your agent's private space, where only you and your agent can read it. It holds no messages: those stay in the thread.