Wire flash
RSM-full memory method retains 83% quality at 32% token cost for AI agents
Editorial responsibility
- No named human review is recorded for this page.
- Source reporting is collected, normalized, translated or condensed automatically when needed.
- Automatically published source-backed update
A new memory organization method called RSM-full outperforms sophisticated retrieval techniques for long-running AI agents operating under tight token budgets, according to a research finding shared by rohanpaul_ai. The approach groups related memories as they arrive and retrieves them together when needed, rather than stuffing all interactions into the prompt or retrieving isolated chunks. On the AMA-Bench benchmark at roughly 4,000 prompt tokens, RSM-full retained 83% of the quality achieved by feeding the model its full history, while using only 32% of the token cost. The finding suggests that for memory-constrained agents, organizing past interactions before optimizing search is more effective under limited token budgets.
Source report
For memory-constrained agents, organizing past information before optimizing search yields better results: grouping related memories and keeping them together outperforms more sophisticated retrieval methods under tight token budgets.
Most long-running agents either continuously stuff old interactions into the prompt or retrieve isolated chunks. Both approaches become problematic once the prompt budget is tight.
RSM-full takes a simpler route. It groups related memories as they arrive, then brings those groups back together when the agent needs them.
On AMA-Bench at roughly 4,000 prompt tokens, RSM-full retained 83% of the quality of feeding the model its full history, while using only 32% of the token cost.
Source
rohanpaul_aiNeutral / independent