Skip to content

Memory, sessions and reminders

Memory

The agent keeps what it knows about you in plain Markdown files, one per topic, in workspace/memory/. The folder is a git repository, so every change has a history. The topics go into the agent's instructions every turn, within a size budget (memory.prompt_budget). When the budget is full, the topics changed longest ago fold to one line.

Memory is written in the conversation, from your own words. After a quiet spell in your private chat (10 minutes by default) or when a session ends, the agent also reads the chat once for facts it missed. You can turn that off with memory.review.enabled: false.

Ended conversations are kept for 90 days, then dropped (memory.archive_days). The facts stay; what goes is the conversation itself, so /resume and the agent's search of past chats no longer find it. Your backups keep it as long as your bucket keeps them.

Every entry passes a check that rejects instructions disguised as facts. Suspicious entries are set aside and you are told. An idea you have for how ganglion itself should change is kept as your idea for its developers, not as a rule for the agent; the check lets it in unless it would let the agent skip a confirmation, gain a permission or send something to someone.

A stored fact is a short line. Ask the agent for details, a write-up or a prompt about something it remembers and it first looks up the chat the fact came from. Ask it for a prompt or a bug report for ganglion's developers and it names the version it runs and asks them to check whether the thing already exists.

/memory lists the topics. /memory kitchen shows one. It works in your private chat only.

Remember, Skip, Undo

When the current conversation has read outside content (a web page, a forwarded message), facts the agent proposes are listed with two buttons, Remember and Skip. Nothing is saved until you press Remember.

After every change the bot sends a note listing it with an Undo button. Undo reverts that change.

Sessions

A conversation is a session. The bot starts a new one after silence, 6 hours by default (agent.session_idle; off never does).

  • /new archives the current session and starts a fresh one.
  • /sessions lists the last 5 earlier sessions.
  • /resume 2 returns to session 2 from that list.

Reminders and scheduled tasks

Ask in your own words:

Remind me every Monday at 9 to water the plants.
Every morning at 7 check the news about Go and send me the three best links.

There are two kinds.

  • A reminder sends fixed text at its time. No model runs.
  • A task runs an instruction as a full unattended assistant turn, at its time.

A schedule is either a cron expression with five fields, or a single date and time in the node's time zone. Check node.tz first: the default is Europe/Kyiv, and it decides what "9 in the morning" means.

A reminder is set at once and shown back to you as a message with a Delete button, so a mistake is one press away. If the session is tainted, the agent asks first, as it does for a task. You confirm a task when it is created. The confirmation lists what it may do: its tools, the web hosts it may fetch, the exact host commands, and who gets the reply. You approve that once. After that, runs need no button presses, and anything outside the approved list is denied with one notice. Tasks count against your daily limits.

/tasks lists them, one message each, with Details, Delete and Run now. When you ask the agent in words to show your reminders, you get the same messages; when it only looks one up for itself, nothing is posted to the chat.

Stopping a reminder

A repeating reminder arrives with a ✅ Done button. Press it and the reminder is deleted at once: the message stays, with a line saying the reminder is off. No model runs. A one-time reminder has no button, since it is over once sent. The button works for a week after the message.

You can also answer in words. Reply to the reminder's message, for example "ordered" or "done, stop reminding", and the agent knows which reminder you mean and deletes it. You get one short line back, such as "🗑 Off: Order a flashlight.", and no question. The same goes for "stop the morning news task" said in plain chat.

If the session is tainted, the agent asks you to confirm the delete first, so a web page or a forwarded message cannot stop your jobs. System jobs, such as the nightly memory audit, and the jobs of modules cannot be deleted this way at all.