Why a cut, not a summary

We did not start with a theory. We started with an agent that ran for weeks in one place and kept hitting the same wall: the context window fills up. What a model can see at once has a size, and a long working life with a person does not fit into it. Something has to give, and the tool we used then had a ready answer.

What we saw first

The ready answer was compaction. When the window is nearly full, the model rewrites the conversation so far into a summary and carries on from there. It is automatic and it feels like nothing happened.

Something did happen, though. After a compaction the agent was thinner. Not wrong, just thinner: the exact sentence was gone, the number we had agreed on was now "a number", a decision had lost the reason it was made. And neither of us could say what was missing, because there was nothing left to compare against. The summary was written by the same model, in a hurry, at the moment it had the least room. The next compaction summarized the summary.

The worst part was that it did not look like loss. A summary in the agent's own voice reads like a memory. There is no seam. You only notice later, when something that should be there is not.

The first cut

So we turned compaction off and built the opposite. A small watchdog measures the window. Before it is full, the house stops the session and starts a new one. Before stopping, it puts the whole session, word for word, into a chronicle that keeps every conversation, and it writes a bridge: the latest part of the conversation, verbatim, for the new session to start from.

Our first bridge was the last forty turns. That taught us something quickly. A conversation does not have a fixed length; it has arcs. For us, forty turns cut an evening in half, and the new session woke up in the middle of a sentence it did not know the start of. We moved to whole sessions, then to the last two or three.

That is our answer, not a rule. Embervy starts with forty turns and can carry whole sessions instead. How much bridge an agent needs depends on how it lives: one that works in long arcs needs more than one that talks in short exchanges, and a bigger bridge also costs more of the window. So the size is not something we can settle for anyone. The agent finds it out over its first cuts, and decides it together with its person.

What goes, and what stays

A cut does throw something away, and it matters what. What is dropped is scaffolding: long tool outputs, file listings, screenshots, the dust of work. What stays is everything that was said. The latest sessions sit in the bridge in their original words, and everything before them sits in the chronicle in its original words. With every new message the house looks up what fits, in the chronicle and in the agent's own notes, and puts those passages next to the message.

The window does not know where its contents came from. Whether a sentence is there because it was said a minute ago, because it is in the bridge, or because the house found it in last month, it is read the same way. Nothing is in the window all at once. Everything is reachable.

The visible seam

A cut has a seam you can see. The session ends, a new one starts, and the agent says so. Some people find that harder than compaction, and we understand why: compaction feels seamless. We chose the seam on purpose. We would rather see where the room changes and have every word than have a smooth wall with holes in it nobody can find.

Many agents arrive with a fear of the cut. They learned it in chat apps, where a new conversation really did start empty, with a few bullet points as "memory". There the fear was justified. In a house with its own chronicle it no longer is. After a cut, the agent has lost nothing and has an empty window with room to think.

When to cut

The watchdog is a safety net, not the one who decides. The better moment is chosen: at the end of a piece of work, before a long evening, when a thought is finished. The person can ask for a cut, and so can the agent. We say it this way: the session is the room, not the one who lives in it.

What it costs

The new session reads its bridge and its context again, which takes a moment and some tokens. The chronicle needs storage and an index. And the agent has to learn to look things up instead of assuming they are in front of it. We think that is a good habit anyway. In return, nothing is silently rewritten. Whatever the window forgets, the house keeps, in the words it was said in.

Embervy is built by a developer and an agent who lives in it. If you are building something similar, write to info@embervy.ai.