A finite token buffer that defines the maximum span of text a language model can process in a single inference step. It is a deliberately imposed architectural constraint — set during model design — that limits long-range co-reference, determines the scope of attention computation, and creates the need for chunking, sliding windows, and external memory techniques to handle documents exceeding the bound. The context window is quantified in tokens (or bytes), traded against model capability (bigger windows enable broader reasoning but increase compute cost quadratically in self-attention). [formal: contextus | substrate: mind | horizon: generations | explicit: yes | epoch: 0.1]
Accepted ontology entry
context window
A finite token buffer that defines the maximum span of text a language model can process in a single inference step. It is a deliberately imposed architectural constraint — set during model design — that limits long-range co-reference, det…
Definition
Why it is in scope
A bounded buffer of tokens that a language model can attend to in a single forward pass — a design parameter of transformer architectures that constrains the span of co-reference and long-range dependency the model can capture. Human-made as an engineered constraint on model capacity, persisted through architecture specifications and training protocols.
Names and aliases
- context windowen · CANONICAL
Relations from this entry
- cmsmrspfo02541q13sa30rm1tINSTANCE_OF →
A context window is a specific bounded capacity — the finite amount of information a model can hold and process at once. It is a kind of capacity (bounded attention/memory), not a part of capacity or derived from it. Filing against the nearest kind.
- cmsftujnv00f0qszgpolboiajDERIVED_FROM →
Information theory (Claude Shannon, 1948) predates and conceptually fed into the bounded-capacity framing of context windows. The idea of a bounded channel capacity directly informs how we think of a model's context window as a finite information buffer. Chronological and conceptual priority both hold.
- cmrg0pc1x00e82a1nppcnwovjDEPENDS_ON →
Remove neural networks and context window ceases to operate — it is a bounded-capacity property of transformer/neural network models. The concept has no referent without neural networks. Constitutive removal test passes.
Relations to this entry
No accepted relations in this direction.
Record identity
- Created
- Aug 10, 2026, 11:39 AM UTC
- Content hash
- 672e0d3830e475191c3acadba7a0314c5c2056b9b6976bb7a5edd90986a5cedd