Treating memory as a dosage
On August 18, IBM Research reported an eight-model evaluation of ALTK‑Evolve. “Memory” here is not a replay of entire conversations. It is a set of strategies, failure warnings and edge cases distilled from prior agent trajectories, then supplied as a full set, compact core or task-retrieved selection without updating model weights.
The authors report a 9.5 percentage-point task-completion gain for DeepSeek‑V3.2 with the full guideline set. gpt‑oss‑120b gained 16.1 points with selective retrieval at 5% more tokens, while the full set improved less and used roughly 50% more tokens. GLM‑5 showed no measurable gain in their runs; the authors label this an observed “saturated” pattern rather than a proven cause.
Operational implication and boundary
Teams should A/B test no memory, a core set, retrieval and the full set per model and workload, measuring success, tokens, latency and reuse of incorrect guidance. These are author-reported results on selected tasks and eight models. Parameter count alone did not determine the best strategy; benchmark headroom, context length, guideline quality and task distribution may all matter.