Open Models Almanac
Methodsinference & servingcontext management

specific method · filed under inference & serving

Discard-all context management

Discard-all is a context management strategy that resets the context by discarding all previous tool-call history.1 It is applied when token usage exceeds a threshold of the context window, extending the token budget so agentic rollouts are not prematurely truncated.2

Trade-off Despite its simplicity, it performs well in both efficiency and scalability, comparable to parallel scaling while using significantly fewer steps.3

How this definition was written

Summarised by glm-5.3-flash from 2 sources; every sentence is checked against the quote it cites.

  1. Discard-all, which resets the context by discarding all previous tool call history — paper
  2. context management employing simple strategies to extend token budgets at test time, when the token usage exceeds 80% of the context window length — DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
  3. Despite its simplicity, Discard-all performs well in both efficiency and scalability, achieving a score of 67.6, comparable to parallel scaling while using significantly fewer steps — paper

Also called Discard-all, discard-all context management strategy, discard-all strategy, Discarding all previous tool-call history.

sources
6
models
4
labs adopt it
2
strongest
used

Documented in

Evidence

7 quotes from 6 sources, one card per model, each opening with its strongest.

How sources treat it
evaluated 3used 4

NVIDIA-Nemotron-3-Ultra-550B-A55B

NVIDIAadopted2 sources

We focus primarily on discard-all context management, where earlier search observations are removed once the interaction history exceeds the context budget

usedunclearin Search teacher for Nemotron 3 UltraNVIDIA
1 more from this model

discard-all context management, where earlier search observations are removed once the interaction history exceeds the context budget

usedinference servingin Nemotron 3 UltraNVIDIA
↑