Remember Less, Preserve the Thread
RunAgent Super Browser treats memory as a state-management problem, not as a larger place to store everything a browser agent has observed. The focus here is what should survive a long task after screenshots, element lists, and old reasoning have become stale. Understanding that distinction helps explain why a compact Ledger can support continuity—and why aggressive compression can also create a new failure if it removes the fact or warning that the next action needs.
A browser agent's live context is a crowded workspace. It can contain a screenshot, a list of elements, a tool response, a session-state block, and the model's prior reasoning for every recent step. Some of that material is indispensable for the next decision. Much of it describes a page that has already changed. The engineering question is not whether the agent can retain more tokens; it is whether the retained tokens still describe the task it is trying to complete.
Why a larger prompt can lose the task
A naive browser loop appends each observation to the next prompt. At first, the history seems helpful. The model can look back at the previous page, recover a label, and remember why it took an action. As steps accumulate, however, old observations compete with current ones. A screenshot from several pages ago may contain the same button in a different state. An element list may preserve a selector that no longer points to the same control. A failed attempt may remain as a long transcript without being converted into a clear warning. The context grows while its signal-to-noise ratio falls.
Super Browser's paper describes a representative twenty-step comparison in which a naive run reaches about 80,000 live-context tokens and falls below 15% prompt-cache hit rate by the end. In the corresponding trace with cognitive eviction, live context reportedly stays between roughly 11,000 and 22,000 tokens and cache hit rate remains above 80%. The authors report about 50% average token saving across their internal evaluation set. These measurements are useful illustrations of the mechanism, not universal constants: task length, provider caching, page complexity, and instrumentation all affect the result.
The failure is not merely financial or computational. Obsolete material can change what the model attends to. If an earlier screenshot shows a filter before it was applied, the model may treat that old visual state as evidence that the filter is still available. If a prior reasoning block explains why a path looked promising, the explanation may continue to influence the next choice even after the path has become a dead end. A longer context window increases capacity, but capacity does not impose a policy about which observation is authoritative.
Super Browser separates live context from structured task memory through a per-task Ledger. The goal remains fixed. A plan and active subgoal can be revised as the task develops. A bounded recent deque holds only the latest three step outcomes, enough to preserve local continuity without keeping the full trajectory in the active prompt. Structured facts record observations with categories, confidence, and the step that produced them. Checkpoints mark meaningful progress, such as reaching a destination or applying a filter.