Long-term memory
Facts and context that persist across conversations, so an agent recalls what happened in earlier sessions instead of asking again.
Memory
Memory
Section titled “Memory”Your agents don’t start from scratch every time. Auteryn gives agents memory so they can carry context across conversations, learn how you like to work, and keep track of what they’ve figured out mid-task.
Three kinds of memory
Section titled “Three kinds of memory”Learned preferences
Over time, an agent picks up how you like things done — tone, format, defaults — and applies them without being reminded.
Working scratchpad
Within a single task, the agent keeps a running set of notes so intermediate findings survive across many steps of a long run.
- Long-term memory spans conversations. Ask your agent about a project you discussed last week and it can pick up the thread.
- Learned preferences span everything the agent does for you — a durable sense of your style that improves the more you work together.
- The scratchpad lives inside one task. On a long, multi-step run, it lets the agent stash and recall its own notes so nothing important gets lost as the work progresses.
Seeing what it remembers
Section titled “Seeing what it remembers”When an agent saves something, a quiet chip appears in the conversation — Saved 2 memories — and when it forgets something, Forgot 1 memory. Both open the Memory panel in the workspace side panel, under Advanced.
The panel has four views — Memories, Learnings, Recently deleted and Me (all agents) — and each entry carries where it came from:
| Badge | What it means |
|---|---|
| Learned from conversation | The agent picked it up from something you said |
| Saved by agent | The agent decided to write it down explicitly |
| Edited by you | You corrected it |
Entries are stamped with the month they were last touched, and anything you told the agent directly as feedback sits at the top of the list.
Learnings is the agent’s own accumulated experience from finishing tasks — patterns it picked up about how work here tends to go, rather than facts about you.
The panel header shows how full the memory is — Standing memory 72 % full · 0 hidden. As it approaches full it turns amber and suggests merging or forgetting entries, because anything past the limit stops being shown to the agent — that’s what hidden counts. The agent sees the same number you do and is told to tidy up rather than quietly losing things. You can save up to 300 entries of your own per agent; what the agent learns on its own is kept trimmed automatically.
During a run, the context ring above the message box includes a Memory row, so you can see how much of the agent’s attention the memory is taking up.
Correcting and forgetting
Section titled “Correcting and forgetting”Editing an entry is just editing text. Forgetting genuinely removes it — the agent stops seeing it on the very next run.
- Deleting has an undo. The confirmation toast offers Undo, and a deleted entry stays recoverable for 30 days.
- Recently deleted in the panel lists what you removed, with Restore on each.
- History on an entry shows how it changed and who changed it, and — where the agent recorded why it believed something — the evidence, under Because you said. History is visible to owners and admins.
Memories the agent learned on its own also fade by themselves. Anything it hasn’t needed in a long time loses weight and eventually drops out, so the list stays about what is current rather than growing forever.
Switching it off
Section titled “Switching it off”Memory is a capability, with three switches on the agent, in the Console under Capabilities → Memory:
| Switch | Effect when off |
|---|---|
| Remember across conversations | The agent starts every conversation fresh |
| Learn from conversations automatically | It only remembers what it’s explicitly asked to |
| Remember what each person prefers | No per-person preferences are kept |
Turning a switch off takes effect from the next run. Nothing is deleted: what the agent already remembers is kept, and turning memory back on restores it. Changing the capability restarts the agent’s sandbox, so finish anything running first.
Just this conversation
Section titled “Just this conversation”For a one-off — a sensitive thread, an experiment, someone else’s problem you’re helping with — open the conversation menu and choose Don’t remember this chat. The message box then shows a Not remembered pill for as long as it’s on.
In that conversation the agent still reads what it already knows, so it’s just as useful — it simply learns nothing new there. Nothing is saved, nothing is extracted, nothing is indexed.
Private conversations
Section titled “Private conversations”Anything an agent learns in a private conversation is kept for you, not for the agent as a whole. A private thread never teaches your colleagues’ agent something they’d see without you meaning it to.
Your memories, across agents
Section titled “Your memories, across agents”Some of what an agent learns is about you rather than about the work — how you like things phrased, your timezone, the name you go by. Those belong to you, not to one agent.
The Me view in the Memory panel shows your own memories across every agent in the organization, with the same edit and forget controls. Forgetting one there forgets it everywhere.
Asking across conversations
Section titled “Asking across conversations”You can ask your agent about things from other conversations — “what did we decide about the pricing page last week?” — and it searches your earlier threads with that agent rather than only what’s in front of it. Answers cite the conversation they came from, with its title and date, so you can go and read the original.
Your private conversations are searchable only by you. Nobody else’s agent can surface them, whatever their role.
Memory on a voice call
Section titled “Memory on a voice call”Talking to your agent by voice in the workspace uses the same memory as chat — the call opens with what the agent already knows about you and the work, and what you say during it is remembered the same way afterwards.
A phone call is different by design: the agent doesn’t assume the caller is you, so a phone conversation doesn’t write to your organization’s memory.
Memory in the file explorer
Section titled “Memory in the file explorer”If your agent has a sandbox, the file explorer shows a .memory folder — a read-only view of what the agent currently remembers, so you can read it alongside the files it is working on. It is a window, not a control: editing memory happens in the Memory panel, where every change has an author and can be undone.
Memory vs. knowledge
Section titled “Memory vs. knowledge”Memory and Knowledge do different jobs:
| Memory | Knowledge | |
|---|---|---|
| What it holds | Context the agent accumulates while working with you | Documents and sources you connect |
| Who fills it | The agent, automatically | You, by connecting sources |
| Best for | Continuity, preferences, in-task notes | Grounding answers in your content |
Use Knowledge when you want the agent to answer from your documents. Memory is the connective tissue that makes an agent feel like it knows you across sessions.