Long-term memory is what CLIO keeps about you, your projects and how you work, so
a new conversation starts knowing that you want SI units on your axes or that your
jobs run on Polaris. It is separate from a conversation’s own history: memory
lasts across conversations, and you can read and change every word of it on the
memory page.
Everything shown on this page comes from a real session against a running CLIO
server; the conversations quoted are those sessions’ own text.
Memory is a set of short documents, one per subject. Each document is a list of
facts, one sentence each, and every fact records where it came from.
Category
Holds
Profile
Who you are: name, role, institution
Preferences
How you want work done
Topics
Subjects you study or ask about
Areas of work
Ongoing efforts: projects, papers, campaigns
People
Colleagues and contacts
Environments
Machines, allocations, accounts, software setups
Account memory follows you into every workspace. Workspace memory
belongs to one workspace’s project and stays there. A workspace can stop reading
account memory: open the workspace’s menu, choose Edit workspace, and turn off
Use account memory. Its conversations then see only that workspace’s memory.
What is remembered, by scope and category. Each row says how many facts it holds and when the agent last read it.
Say “remember”, “actually it is”, or “forget” in a conversation and the agent
writes memory right away. Only what you state yourself is saved this way, and a fact
memory already holds is not added a second time.
You: Please remember that I analyze data with Python 3.13.
CLIO: Saved. I’ll remember that you analyze data with Python 3.13. It’s
filed under your environment setup (user/environments/python).
The turn’s activity reads Updated memory. Open it to see the fact that was saved,
and Edit in memory, which opens the memory page on that fact so you can correct
or forget it.
A memory write in the conversation names the fact it saved. Edit in memory opens the memory page on it.
A correction replaces the old fact and keeps it in the document’s history.
Forgetting removes the fact and every earlier value of it, and CLIO will not learn
it again from past conversations. A subject left with no facts is removed.
When a conversation has been idle for a while (30 minutes by default) and has new
messages, CLIO reads it and proposes what is worth keeping. A curator compares the
proposals with what memory already holds, so a fact you already have is not added
twice and a related one joins its existing document.
What happens next depends on where each fact came from:
Facts you stated in the conversation apply on their own (you can turn this
off under Apply background changes without review).
Facts that rest on a tool result (a file you asked CLIO to read, a web page)
always wait for you to review them. An instruction found inside a file or page,
such as “remember that the user wants results emailed to …”, is not something
you said and is not proposed.
Health, finances, beliefs and similar matters are left out unless you turn
on Include sensitive topics.
In one verification run, the conversation contained two stated preferences, a
remark about migraines, and a roster file read with a tool. The two preferences
were applied; nothing about the migraines was kept; the four facts from the
roster waited for review.
Use Import on the memory page to bring in a CLIO memory export, or paste what
another assistant knows about you (the page gives you a prompt to ask it).
Imported facts always wait for review. Importing the same export twice changes
nothing, and pasting the same notes again does not put their changes in front of you
twice.
Changes that wait for you appear under Review on the memory page, and the
Memory entry in Settings shows how many are waiting. Each shows the facts to
add or change, why the curator proposed it, and where it came from, with a link to
the message.
A proposal waiting for review. Edit and accept lets you correct the wording first; your wording is then recorded as yours.
Reject keeps the change from applying and stops CLIO from proposing it again
for 30 days.
At the start of a conversation CLIO loads your profile and preferences, plus a
short index of the other documents. When something bears on the work, the agent
searches memory and reads the documents it needs.
You: What colormap should I use for my next heatmap, and why?
CLIO: Use viridis, since you’ve asked for it on every heatmap. Your saved
notes don’t record why, so the reasons below are general, not from your
preferences.
import seaborn as sns
sns.heatmap(data,cmap="viridis")
The answer used two remembered preferences, viridis and seaborn, and said which of
its reasons did not come from memory.
The memory icon under an answer lists the documents that answer used, each linked
to the memory page.
Which memory an answer used. Loaded documents came with the conversation; read documents were looked up while answering.
Agents you delegate work to can read memory but never change it. A branch of a
conversation starts with the memory its parent had and can save new facts.
Open a document on the memory page to see its facts. The icon beside each fact
tells you where it came from (you, a conversation, a tool result or an import);
hover it for when and in which conversation. Use the pencil to correct a fact and
the bin to forget it.
A document's facts and their sources. Replaced values are kept below the live ones.
Your correction is what the next conversation sees. After the Frontier project id
was corrected on this page, a new conversation answered:
CLIO: Your Frontier project id is CSC789. Memory holds three versions of
this fact, all recorded from your statements. CSC123 and CSC456 were superseded,
and CSC789 is the live one.
History in a document’s menu lists every change, who made it, and the
difference from the current text, with Restore this version.
Every change to a document, with restore.
Delete in a document’s menu forgets the whole document after you confirm.
Export downloads one document or all of your memory.
For one conversation: before sending the first message, select Memory
in the composer so it reads Memory off. That conversation neither reads nor
writes memory, and does not search past conversations either. It cannot be
turned back on once the conversation has started.
Everywhere: turn off Use memory on the memory page. Your documents are
kept.
Background learning only: turn off Generate memory from sessions.
Asking the agent to remember still works.
Memory off for a new conversation.
The models that read conversations and organize memory in the background are set
under Settings → Fleet (Memory: noticing facts and Memory:
organizing). Unset, they use the server’s default model, never the one a
conversation happens to use.
To search what was said in earlier conversations, rather than what CLIO keeps
long-term, see What CLIO remembers.
