Adaptive memory

We didn’t want memory that silently hoards everything. We wanted memory that gets better with use—and that you can see, edit, and turn off.

So we built two layers: manual memory you curate yourself, and adaptive memory that watches your conversations and suggests what to keep—only when it looks like you’re sharing something about you (preferences, facts, “remember this”), not when you’re asking for help or venting.

In the Memories drawer you get a clear list of short, editable memories (up to 50 words each) and a counter (e.g. 12 / 250) so you always know your limit.

A single toggle turns memory collection on or off. When it’s off, we stop learning from new chats; what’s already saved is still there and still used.

After an assistant reply, a small status badge appears while we check whether that exchange should change your memory; it then turns into “Memory Created”, “Memory Updated”, or “Memory Updates” so you know what happened without opening the drawer.

For big, sweeping updates we pause and ask: “Your message would update your memory and profile. Would you like to proceed?” You can confirm or cancel—you keep a veto on big changes.

When collection is on, we run a two‑step process after each reply: “Should we remember anything from this?” (only if it looks like facts or preferences about you), then “What should we add, change, or remove?”

We create new memories, update notes on existing ones, delete what you contradicted, and consolidate so one topic stays one memory.

Memories are scoped by context: global for default chats, project‑specific when you’re in a project, and separate for personas so “you” and “the persona” never get mixed.