Agent Memory Across Steps
Context Limits in Practice, back in Prompt Engineering, showed a human conversation slowly filling up a fixed-size window until something had to be dropped. An agent's scratchpad fills up the exact same way, only faster. Every Thought, every Action, every Observation it receives back from a tool all sit in the same one context window, right alongside whatever the user originally asked for.
Predict before you look
This looks like the earlier conversation widget, but every non-system turn here came from the agent itself, reasoning, acting, and receiving results, not from a person typing. Before you try the tighter limits, do you expect an agent's own scratchpad to fill up faster or slower than a normal back-and-forth conversation?
Tighten the limit and the earliest Observations disappear first, the exact same rule as the human conversation, just applied to the agent's own research history instead of someone else's messages. That has a real consequence here that a plain chat doesn't, if an early Observation gets dropped before the agent writes its final answer, the fact it found is gone, not summarized, not remembered elsewhere, gone, and the final answer has to be written without it.
This is why real agent systems don't just let a scratchpad grow forever. Some periodically summarize older steps into a shorter form before the window fills, some cap how many tool calls a single loop is allowed to make, some write important facts somewhere outside the context window entirely and search for them again on demand instead of holding everything in the live window at once. None of that is a new mechanism, it's the same finite context window this whole site keeps returning to, just now the thing filling it up is the agent's own work instead of a conversation with a person.
What happens if an agent's Observation from three steps ago gets dropped before it writes its final answer?