How Lunar remembers

Most AI chats have the memory of a goldfish: close the tab and everything is gone. Lunar is built the other way around — it is designed to accumulate. Memory works on three levels, and once you understand them you can use them deliberately.

Your Space keeps the long arc

When you create a Space, you give it a purpose and instructions: what this semester is about, how you want Lunar to teach you, what to prioritize. You write this once, and Lunar keeps it present in every class of that Space — so each conversation starts already aligned with your goal, without re-explaining your situation.

On top of that, each Space builds accumulated memory: a compact, auto-maintained record of the long arc of your semester, plus the topics already covered in its concept graph. This is why week-three classes are aware of week-one doubts, even in a brand-new chat.

And when a Space holds many classes, Lunar does not drag all of them into every message: it automatically picks the previous classes most relevant to what you are asking right now and brings only those in. Your history works for you without burying the answer.

[!TIP] Write Space instructions as if briefing a tutor who will teach you all semester: your level, your goal, your preferences. The more precise you are, the better every class inherits it.

Every chat keeps a living session context

While you talk, Lunar maintains a session context for the current chat. It tracks four things:

• Objective — what this session is trying to achieve.
• Covered — the ground you have already gone through.
• Open questions — doubts that appeared but were not fully resolved.
• Next step — where the conversation should pick up.

Alongside those notes, Lunar also reads back its own session memory — a recent summary of the current conversation plus the last exchanges between turns. Practical consequences: you can take a break and return without repeating yourself, follow-up questions build on the whole session instead of the last message, and you never have to re-explain what you said two turns ago.

Your learner profile shapes every answer

Lunar keeps a picture of who you are as a learner: your level, your goals, your language and the study preferences it observes over time. It uses that profile to tune explanations — vocabulary, depth, the kind of examples that land — automatically, in every model: Classic, Creative and Mister all adapt the same way. Switching models never resets what Lunar knows about how you learn.

You can see what Lunar remembers about you

That learner profile is not a black box: it is visible in Settings → Memory, together with the evidence that originated each datum and its confidence level. It is read-only on purpose: memory is corrected from the conversation itself — saying "actually, that's not right" is the most valuable correction there is — not by hand-editing it, so every change is backed by something you actually said or did.

On top of that, memory stays fresh on its own: data that go a long time without confirmation fade from daily work until something reconfirms them, and when the profile grows too large, Lunar compacts it automatically keeping what matters. Remembering a lot does not mean dragging everything forever.

Smart titles, no renaming required

Every conversation is titled automatically from its content as you work, so your sidebar stays useful without anyone naming anything. And if you do rename a chat by hand, that name is yours to keep — Lunar never overwrites a manual title.

Microchats remember too

Each microchat generates its own context file when you close it, and the main chat takes relevant microchats into account in later answers — so a doubt you resolved on the side is not lost when you return to the main thread.

What this means in practice

• Refer back to past classes. Ask things like "how does this connect to what we saw last class about sampling?" — Lunar can tie new material to earlier sessions in the same Space.
• New chats start warm. A class in an existing Space does not begin from zero: it knows the covered topics and the open threads.
• Your open questions become a plan. The "next step" tracking means revisiting a topic continues where it stopped, instead of starting over.
• Continuity survives branching. Branches inherit a context seed from their parent, so side paths start coherent.
• Explanations come pre-tuned. Your learner profile and the session memory work in the background on every model, so answers start at your level rather than a generic one.

[!IMPORTANT] To get the most out of memory, keep related work inside the same Space and continue threads instead of opening blank chats. If the memory ever feels off, correct it: edit the microchat's context file or refine your Space instructions — your words take precedence.