Lunar's memory, in three levels

Lunar does not treat all memory as one bucket. A personal preference is not the same as a class objective, and a class is not the same as an entire Space.

Memory about you

Global memory stores what Lunar notices about you and can use across conversations:

• how you like explanations to be delivered;
• the tone and level of detail that work for you;
• interests, personal context, and topics you often explore;
• facts you explicitly stated.

It does not store the goal of one class or whether you mastered subnetting in one conversation. Those belong elsewhere.

You can view and edit it in Settings → Memory. Facts you state explicitly take priority over a later inference: if you say you prefer examples, Lunar should not replace that with a contradictory guess unless you change it.

Memory of each chat

Each conversation can have a living context file. It stores what only makes sense for that session:

• the class objective;
• topics covered;
• goals for that study session;
• knowledge state for those topics;
• open questions and the next useful piece of work.

So “prepare for the networking exam” does not become a permanent fact about you. It stays in the chat where it belongs.

Memory of a Space

A Space groups several conversations — for example, a subject or project — and has two layers:

• Space-wide memory: purpose, instructions, and connections that help across all its classes.
• Per-chat memory: the objective, topics, goals, and knowledge of each individual conversation inside the Space.

The first helps a new chat understand the shared context. The second lets you open one class without mixing it with every other class.

What Lunar sees on each turn

Lunar receives a compact selection from each level:

1. memory about you, to adjust how it explains;
2. Space memory, to keep continuity across classes;
3. the current chat's context file, to know what is being studied now.

Memory does not replace Lunar's core instructions. It supplies context. If a system instruction says how the tutor should behave, a personal preference does not override it: it adjusts examples, tone, and depth within that frame.