A study of eight models reveals that agentic memory is not a universal feature but a dose that must be calibrated to the specific model's capability. The ALTK-Evolve framework distills guidelines from an agent's past trajectories and injects them at inference time without weight updates, showing that optimal memory strategies vary significantly across model tiers.
- Strong models with headroom benefit most from a full guideline set, as seen with DeepSeek-V3.2 gaining +9.5pp task completion.
- Weaker models perform best with curated retrieval of a compact core plus per-task guidelines, allowing gpt-oss-120b to gain +16.1pp at only +5% token cost.
- Saturated models showed no measurable gain, suggesting they have reached their performance ceiling on the tested tasks.
- Curated retrieval offers both higher accuracy and lower cost for weaker models compared to injecting the full guideline set.
The authors conclude that memory should be calibrated rather than merely accumulated, as giving an agent more experience than it can use yields no benefit. They recommend compact cores for weak models and prompt caching for strong models to keep costs manageable in production.