IBM Research and Hugging Face have released findings suggesting that the amount of contextual memory an AI agent needs depends fundamentally on the model's capability level, challenging the assumption that more memory always improves performance. Across eight models—ranging from 30B dense models to frontier systems—researchers found that strong models with computational headroom benefit from complete guideline sets, while smaller models perform better with selective, task-relevant retrieval, and already-saturated models show minimal improvement regardless of memory provision. The team tested their ALTK-Evolve framework on 585 multi-step tasks across nine simulated applications, measuring task completion and scenario success rates. A striking result: gpt-oss-120b gained 16.1 percentage points in task completion using curated retrieval while increasing token usage by only 5%—suggesting that calibrated memory approaches can outperform blanket memory injection on both effectiveness and cost. The researchers note that memory calibration appears influenced by benchmark headroom, context-window size, model architecture, and task distribution, not merely parameter count. The approach carries immediate implications for enterprises deploying agentic systems. Rather than applying uniform memory strategies across models, teams can now tune guidance based on their specific model's characteristics. Because ALTK-Evolve requires no weight updates or human annotation, it is portable across architectures and affordable to implement in production, particularly when combined with prompt caching.