How agents tidy their memory overnight

Agents remember things about you and your workspace as you work together — your preferences, decisions you've made, ongoing commitments, and useful context. Over time, that memory can pile up: the same fact noted twice, an old preference that's since changed, a detail that no longer matters.

To keep that from turning into clutter, your assistant does a little housekeeping while you sleep.

What happens overnight#

Once a night, during quiet hours, your assistant re-reads what it has learned and tidies it up. It will:

• Merge duplicates — fold two notes that say the same thing into one.
• Resolve contradictions — when something changed (you moved cities, a deadline shifted), keep the newer, more reliable version and retire the old one.
• Tighten verbose notes — shorten rambling entries without losing the facts.
• Retire stale notes — set aside low-value details it hasn't needed in a long time.

This runs in the background. It never interrupts you, and you don't have to do anything to enable it.

Small instruction fixes, too#

Sometimes the day's conversations clearly overtake a single line in your agent's instructions — for example, the instructions say to post the daily summary in one channel, but you've since asked it to use another. On the nightly pass, your assistant can propose a small, one-line fix to that exact line.

These instruction fixes are held for your approval by default; switch on auto-apply if you'd rather they take effect immediately. Either way, every fix keeps the previous version and can be reverted in one click. (Bigger, ground-up rewrites of an agent's instructions are a separate feature, covered in How agents improve over time.)

Keeping an agent's description honest#

As a teammate takes on more work, its short description — the one-line "what this teammate is for" that helps Monad and your team route questions to it — can fall behind what it actually does. On the same nightly pass, if an agent has clearly grown beyond its description, your assistant can propose a tightened, more accurate one (without letting it stray into another teammate's area). Like the instruction fixes, a description change is proposed for your approval by default and can auto-apply per agent once you trust it — and, also like them, every change keeps the previous version on record so it can be reverted.

Turning it on or off#

Go to Settings → Agents → Optimization → Memory & nightly fixes (workspace admins) to turn the nightly tidy-up on or off per agent and to choose whether instruction fixes auto-apply.

The nightly run is the master switch for overnight improvements: if you turn it off for an agent, that agent's periodic instruction rewrites (under Prompt optimization) pause as well, since they run as part of the same nightly pass.

Nothing is lost#

When your assistant retires a memory, it doesn't delete it — it just sets it aside. Retired memories stop showing up in conversations, but they're kept on record and can be brought back if needed. Every change is logged with a short reason, so there's always a trail explaining what changed and why.

The tidying is deliberately gentle: it makes a handful of careful edits rather than rewriting everything, and an unusually large nightly change is flagged for review. The goal is a memory that gets cleaner and more accurate over time — not one that forgets things it should keep.

Privacy stays intact#

Tidying happens one memory area at a time, and the areas never mix:

• What your assistant knows about you is tidied on its own and never blended into shared, workspace-wide notes.
• Shared notes for a channel or thread are tidied separately, visible only to the people who can already see that space.

So this housekeeping never moves a private detail into a place where someone else could see it.