Paper VISTA: Let Tool Agents "See Their Own Context" — Self-Managed Memory Without Training
arXiv:2606.30005 introduces VISTA: long-horizon tool agents are bottlenecked by context growth, yet models are "proprioceptively blind" to their own context. VISTA is a training-free, model-agnostic interface exposing working memory as addressable blocks so the model itself makes keep-or-drop calls.
Problem: models are "proprioceptively blind" to their context
Long-horizon tool agents are bottlenecked as context grows toward the window limit. Recent systems make context management agent- or system-controlled, but either learn a compression policy that discards evidence, or manage it in a layer the agent never sees. This paper (arXiv:2606.30005) points to a more basic gap: frontier models are "proprioceptively blind" to their own context — from the prompt alone they can't see how large, how old, or how often each block was used, exactly the signals a keep-or-drop decision needs.
Hypothesis: the ability exists; the interface is missing
The core hypothesis: competent context management is already latent in capable models; what's missing is not a learned policy but an interface that surfaces this state.
Method: VISTA — visible internal state
VISTA (Visible Internal State for Tool Agents) is a training-free, model-agnostic layer that:
- represents working memory as typed, addressable blocks;
- surfaces a runtime dashboard of per-block token usage, recency and access history;
- archives blocks as recoverable, full-fidelity payloads (dropped but retrievable — no lost evidence).
On LOCA-Bench, BrowseComp-Plus and GAIA, the same untrained interface lets models make better context trade-offs on their own. VISTA's contribution: reducing "context management" from "train another policy" to "give the model a visible state dashboard," letting it use existing abilities to manage memory — a useful reference for engineering long-horizon agents.