Engram study: RL training teaches agents to use notes, adding 7-20pp accuracy at 5-30% of memory system cost

sarahcat21 · x · 2026-11-05

Engram published a study on how text memory (notes, wikis, skill files) and RL training complement each other:

With fixed high-quality notes and an optimized prompt, RL training yields 7-20pp accuracy gains across benchmarks, beating both untrained models and frontier models given the same notes
Training costs roughly 5-30% as much as building a good text memory system, and amortizes within days via cheaper inference
Trained models are robust to note drift: trained once on initial notes, they still beat untrained ones after 23M tokens of edits, learning general note-using strategies rather than overfitting

Takeaway: training and text memory multiply, making agents smarter, cheaper, and more robust.

More from coding & agent