gpt-oss-120b gained 16.1pp task completion with 5% more tokens

Hugging Face reported that gpt-oss-120b achieved a 16.1 percentage point task completion gain with only 5% more tokens when using selective memory retrieval instead of full guideline injection, while dumping the full memory set degraded performance and inflated token costs by 50%. The finding suggests that mid-tier models benefit from restricted, retrieved subsets of distilled guidelines, whereas frontier models like DeepSeek-V3.2 can handle full-set memory injection.

Hugging Face https://huggingface.co/blog/ibm-research/altk-evolve-hmm gpt-oss-120b gained 16.1pp task completion with 5% more tokens Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. Tailoring agentic memory via selective retrieval yields a +16.1 percentage point task completion boost for mid-tier models like gpt-oss-120b at just +5% token overhead, whereas dumping the full memory set degrades their performance and inflates token costs by 50%. This means you cannot use a one-size-fits-all context injection strategy in production; you must restrict smaller models to a tight, retrieved subset of distilled guidelines to prevent cognitive drowning while reserving massive, full-set memory injection only for frontier models like DeepSeek-V3.2. 16.1pp task-completion gain on gpt-oss-120b with only +5% token cost when using selective memory retrieval instead of full guideline injection. This means you can ship stronger agents on mid-tier models without blowing up inference budgets—just swap static prompts for a lightweight retrieval layer that serves only the most relevant past lessons per task.