Basic idea of LLM context

What the model can actually see

A model answers from the input it receives on this request. A transcript stored elsewhere does not become active context until the harness includes or retrieves it.

VISUAL STUDY / STORAGE ≠ ATTENTION

Available later. Visible now.

STORED HISTORY

Everything retained.

m00001Task intent

m00002Earlier file read

m00003Past decision

m00004Consumed test log

m00005Current question

Persistence keeps sources available.
It does not put them into attention.

select or retrieve

REQUEST INPUT

What the model sees.

Task intentincluded

Relevant past decisionretrieved

Current questionincluded

An absent source cannot directly support the answer until it is brought into this view.

THE NOTEBOOK

Concise learning notes · source-grounded mechanisms

An LLM itself does not inherently maintain long-term conversational memory.

Context is the information available to the model during the current inference/request.

Therefore:

conversation history ≠ active context

A long conversation may contain much more information than what is currently placed inside the model's context window.

CHECK YOUR INTUITION

An old message is saved on disk but absent from this request. Can the model directly attend to it?

Choose an answer to reveal the reasoning.

FOLLOW THE SOURCE

Reviewed source: billion-context 0.1.180 / 03d27f9. The notes are Lawson’s learning interpretation of the reviewed repository.

Living preprint · architecture and doctrine ↗src/compress-tool.ts ↗