DeepSeek-V3.2 and GLM-5 differ on memory — here’s why
huggingface research shows agentic memory requires calibration per model, with deepseek-v3.2 gaining +9.5pp accuracy from full guidelines while smaller models failed with the same load.
huggingface researchers released a study analyzing how 'agentic memory' — distilled guidelines from past agent trajectories — impacts performance across eight different llms. contrary to the assumption that more memory always helps, the results show that memory is a dose requiring calibration: strong models like deepseek-v3.2 gained +9.5pp task completion with full guideline sets, while weaker models like the 117b gpt-oss-120b preferred a selective retrieval method to avoid context noise. saturated models like glm-5 showed no measurable gains regardless of memory injection, likely due to performance ceilings or irrelevant guidance. because this method alters the agent's context rather than model weights, it offers a cheap, portable way to improve agents, provided developers tailor the memory size to the specific model architecture.