ALTK-Evolve implements agentic memory by distilling reusable guidelines from an agent’s past trajectories and re‑injecting them at inference time, without updating model weights or requiring human annotation. The approach treats memory as a dosage that must be matched to the model’s capability rather than a binary on/off switch.
Evaluating eight models ranging from a 30B dense transformer to a 671B MoE system revealed three dosage patterns. Strong models with unused capacity benefit from receiving the complete guideline set—DeepSeek‑V3.2 (671B MoE) showed a +9.5pp rise in task completion under this condition. Weaker models achieve higher accuracy with a compact core plus per‑task retrieval, while saturated models exhibit no measurable gain. Curated retrieval proved both accurate and economical; gpt‑oss‑120b gained +16.1pp completion at only a +5% token overhead, and prompt caching keeps the full set affordable in production.
- ALTK-Evolve: guideline distillation, inference‑time injection, no weight updates, no human labels.
- Memory dosage: full set for high‑capacity models, core + retrieval for mid‑tier, none for saturated.
- Benchmark gains: DeepSeek‑V3.2 +9.5pp, gpt‑oss‑120b +16.1pp (+5% tokens).
- Prompt caching reduces cost of large guideline injections.
Why this matters
Source facts indicate that aligning memory volume with model headroom yields measurable task‑completion gains without extra training. The inference is that enterprises can dynamically allocate memory budgets according to model size and workload, avoiding over‑provisioning while still capturing the accuracy benefits of agentic memory.