Recall · Palette

Anatomy

When an answer depends on something the model remembers from before: a preference, a reference figure, a house style.
Whenever the model learns on its own. A learned memory people can’t see is one they can’t correct.
Not for facts from a document the model just read. Those are sources: show them with Aperture or Trace.

Action	Result	Rewritten
Open a marker	The memory, where it came from, and the parts of the answer it shaped	Nothing
Correct	The memory changes and becomes one you told: ink, not pigment	Only the parts it shaped
Forget	The memory goes and the answer falls back to the default	Only the parts it shaped
Accept	Rewritten parts settle into still pigment	Nothing
Escape	The memory closes	Nothing

A correction is something you told the model, so its marker turns from pigment to ink. The model may learn; only a person confirms.

Mark a memory where it acts, not only in a settings page. People notice a memory when they see its effect.
Say where each memory came from: “You said this on 2 Oct” or “Learned from three conversations in September”.
Tell learned from told: pigment for what the model inferred, ink for what the person said.
Rewrite only what depended on the memory. Leave the rest as accepted.

DoShow the memory where it shaped the answer, with Correct and Forget beside it.

Don’tDon’t update memory silently. People can’t correct what they can’t see.

Each marker is a button named for its memory: “Memory 2: Use the June valuation as the reference.”
Told and learned differ in shape and in words, not only in colour.
Corrections are a labelled group with pressed states; Escape closes the memory.
A rewrite announces how many parts of the answer changed.