Automatic Memory Management asks before it changes how the AI behaves

Most AI memory is a quiet profile. The app watches what you say, builds a picture of you somewhere you can't see, and slowly changes how it talks to you — without showing you what it learned or when the change took effect. That's convenient right up until the day the assistant answers differently and you can't find the line that caused it.

just4o.chat's memory works the other way around. Every memory is an item you can open, and any change to how the AI behaves across your chats waits for your approval.

The feature used to be called Adaptive Memory. It's now Automatic Memory Management — automatic because a background agent does the upkeep, managed because you stay the one in control.

Every memory is an item you can read, edit, or delete

A background agent creates, updates, consolidates, and deletes memories after your messages. None of that happens in the dark. Each memory is a discrete entry in Settings → Memories: open it, rewrite the text, attach a note, or delete it. Entries run up to 200 words, and how many you can keep per scope follows your plan — 15 on Explorer, 30 on Builder, 75 on Studio, 100 on Aurora, 200 on Pro.

Every write has to be earned. The agent can only save a memory when it can cite the user messages that justify it, so each entry traces back to something you actually said rather than something the model guessed about you.

Global and project memory stay separate

Project chats use project memories. Global chats use global memories. What you tell the assistant inside one project doesn't bleed into unrelated conversations, so a memory you set up for a client's work won't quietly reshape your personal chats.

When it wants to change how the AI behaves, it asks first

This is the bigger change. Saving a fact about you is one thing. Editing your Custom Instructions — the durable rules for how the assistant should reason, format, and speak across every chat — carries more weight, so it's gated.

When the agent proposes a Custom Instructions edit, you get a Review customization panel: your current instructions on one side, the proposed version on the other, both editable. The prompt is plain: "Approve or edit the proposed customization before it changes future replies." Nothing takes effect until you choose Apply customization. If you'd rather not be asked each time, there's an "Auto-approve future customization edits" toggle — and it's off by default.

A heads-up when a lot changes at once

If the agent makes several memory changes in a single batch, a notice surfaces instead of letting them pile up silently: "Automatic Memory Management made several memory changes. You can review the details or leave them as-is." You can Review changes or pick Not now.

Two switches, because retrieval and writing are different decisions

Memory has two independent toggles. One controls whether memories are pulled into your chats at all. The other controls whether the background agent is allowed to write. Turn writing off and keep retrieval on, and memory becomes read-only — it still informs your chats, but nothing changes unless you change it. The toggle says exactly that: "Off — memory can still be retrieved when Memory is on, but changes stay manual-only."

It runs in the background, on a small fast model

The agent runs after your message, in parallel with the reply, so it never slows down the model you actually chose. The bookkeeping is handled by a lightweight Gemini Flash-class model rather than the one writing your answer. And it stays lean on the way back in: instead of stuffing your entire memory into every prompt, it pulls roughly the top fifteen entries most relevant to the message in front of you.

Automatic, but managed

Model choice at just4o only matters if the work around the model survives the switch — your files, projects, context, and memory should stay put whether you're asking Claude, GPT, Gemini, Grok, or Qwen. Memory earns its place in that set only if you can see it and steer it. So you can: read every entry, approve the changes that matter, and switch any of it off whenever you want.