Give your agent a 'proprioceptive dashboard' of its own context — Gemini-3-Flash jumped 22.7% → 50.7%

arXiv 2606.30005·high signal

VISTA (arXiv 2606.30005) makes working memory a set of typed, addressable blocks and surfaces a runtime dashboard showing each block's token size, recency, and access history, with archived blocks recoverable at full fidelity. Models are otherwise blind to their own context state, so they can't make good keep-or-drop decisions. Training-free and model-agnostic — it lifted Gemini-3-Flash from 22.7% to 50.7% on LOCA-Bench and reached 58.0% on BrowseComp-Plus. For builders: stop auto-compacting blindly; expose usage metadata and let the agent evict its own blocks.