Memory
How Suzent remembers you across conversations, where that memory lives, and how to change it.
Sovereign continuity. Memory is where a sovereign AI agent keeps its self. Suzent's memory is Markdown on your own disk, not rows in a vendor's database, so you can read it, edit it, track it in Git, and carry it to another machine. That is why switching models does not reset the agent that knows your work. See what makes an agent sovereign.
Suzent remembers facts you've shared, your preferences, and context from past conversations. Memory is on by default and needs no setup.
How it works
During a conversation. Every few exchanges, or once a chat has been quiet for about ten minutes, Suzent reads what was said since it last looked and picks out anything worth keeping, such as a preference, a project detail, or a fact about you. It adds those to that day's log. It saves facts, not whole conversations. Reading several exchanges at once means that if you change your mind partway through, only where you ended up is saved.
A fact about you is only kept when you said it yourself; something the agent said or guessed is not. Each saved fact carries the words it came from, so you can see why it was kept. Facts the agent read in a file or on a web page are kept too, with where they were read. Scheduled tasks are left out of facts about you, and so is anything that only matters today. Facts are saved in the language you write in, so if you chat in Chinese, your memory is in Chinese too.
Lessons come from a separate check. When something goes wrong (a command fails, or you tell the agent it got something wrong) and the same exchanges then show what fixed it, the agent saves what to do next time and where it applies, such as a repository or a tool. A failure that was never fixed, or that will not happen again, like an expired login, teaches nothing and is not saved. This check also covers scheduled tasks. When a lesson already in your notebook comes up again, it is counted rather than saved twice; once a lesson about a repository has come up three times, the organizer asks whether to add it to that repository's agent instructions.
Overnight. A background pass, the organizer, reads the new logs and folds them into tidy pages: one page per topic, duplicates merged, contradictions resolved. Saving stays fast during the day, and the tidying happens later. The Organize tab on the Memory page shows how far it has got and what it did last time, and has an Organize now button.
There are two kinds of memory:
| What it holds | Where | |
|---|---|---|
| Conversation memory | Facts about you, gathered automatically | /shared/memory/ |
| Notebook | Knowledge pages the agent researches and writes | /mnt/notebook/ |
The notebook is an Obsidian-compatible vault. See Notebook.
Where your memory lives
/shared/memory/
MEMORY.md # The map of your memory the agent always sees
archive/
2026-08-24.md # One log per day
persona.md, user.md # Editable profile notesThese files are the real memory. Open them in any text editor. The search index is rebuilt from them, so a line you delete here is gone.
MEMORY.md is a map
MEMORY.md is loaded into every conversation. It is a short list of links,
one per notebook page, each with a few words on what the page holds:
- [[3_Personal/Preferences]] - how you like answers, tools, and formatting
- [[2_Wiki/Experience/Release checks]] - what went wrong in past releasesThe facts themselves live on those pages, so the map stays small (30 lines by default). If the file is missing, Suzent drafts one from your page titles; after that it is yours, and the organizer keeps its links current.
user.md is for what the agent needs in every conversation, such as your name,
your language and how to address you. Schedules and the state of a project do
not belong there, and other facts about you go on a 3_Personal/ page.
Managing what's remembered
The Memory panel in the app lets you browse, search, and delete what has been saved.
To see what memory did in the chat you are in, open the overview in the right sidebar and expand Memory. It lists the memories the last reply drew on, the latest facts saved from this chat (corrections are marked Updated), and whether organizing is caught up. Open memory takes you to the full panel.
- To forget something, delete it in the panel. It disappears from search for good, even after a rebuild. The original daily log keeps the line as history.
- To correct something, just say the corrected version in a conversation. The organizer replaces the old fact, and keeps the old value under a History heading if it was true once.
- When something is wrong, say so ("that's not right, I never lived in Basel"). Suzent hides it from search straight away and will not save it again.
- When the organizer is unsure, for example two facts disagree, it writes a question instead of guessing. The agent may ask you a few of these when they come up naturally, and your answer settles them.
How memory is tidied
Roughly once a day, the organizer reads logs it hasn't processed and, for each fact, does one of four things:
- Confirm: the page already says this, so it raises the page's
confirmedcount. Facts you repeat often rank higher in search. - Sharpen: the page says something vaguer, so it improves the wording.
- Supersede: the page is now wrong, so it corrects it.
- Add: the fact is new, so it writes it down.
Each page says what kind it is: facts about you, lessons from past work, or
general knowledge. The organizer files each kind where your notebook's schema.md
says, or beside existing pages of the same kind, or in the closest folder you
already have. In a new notebook that means 3_Personal/ for facts about you,
2_Wiki/Experience/ for lessons, and 2_Wiki/ for knowledge. Your folders
don't have to match: the organizer never edits schema.md and never adds numbered
top-level folders. To steer it, say where each kind goes in schema.md.
The organizer never touches today's log, and an interrupted run is retried rather than
skipped. Everything it changed is listed in the notebook's log.md.
Facts expire, gently. Each fact about you gets a review date based on its kind: never for identity, a year for preferences, six months for technical setup, three months for goals, three weeks for passing context. After that date, the organizer re-checks it against recent conversations. Expiry never deletes anything; it only lowers how strongly an unconfirmed fact shows up in search.
When nothing new needs consolidating, the organizer audits the notebook instead, looking for contradictions, broken links, orphaned pages, and stale knowledge.
It tells Regent what deserves attention. When a run turns up something worth knowing now, the organizer sends Regent one short message: several lessons about one tool that could become a skill, the same failure coming back, a project that has gone quiet, or something notable it saw. It can't message anyone else, and most runs send nothing. Regent decides whether to tell you, and if you agree, it can start a session to act on it.
Settings
Turn memory on or off in Settings → Memory System. Its models are set in Settings → Model Roles: the organizer uses the Memory organizer role, picking out facts uses Memory Extraction, and memory search uses Embedding (see Model roles).
How often the organizer runs, where the notebook lives, and other tuning are in Advanced memory settings.
Checking and tidying your memory files
Two commands help when memory has grown untidy:
suzent memory lintsuzent memory cleanuplint lists what does not follow the rules above: a MEMORY.md that holds
facts instead of links, an overlong user.md, oversized pages. cleanup writes
suggested fixed versions next to your memory for you to read; it never changes
your files. See Command line.