Context and memory
Nothing gets dropped to make room.
Nearly every product handles a full context window the same way: silently throw away the oldest messages. That's how an agent contradicts a decision you made an hour ago.
Long sessions become chapters
The full conversation is stored permanently. What changes is the view: older stretches condense into chapters, and as a session grows, chapters condense again.
The agent carries weeks of history as a structured outline plus the recent conversation verbatim. When it needs the exact command or the exact number, it expands that chapter back to the original messages and reads them.
Summaries are how it navigates. Raw history is what it trusts.
You get the agent's view
Chapter markers sit in the scrollback and a minimap you can scrub runs alongside it.
Finding the moment you half-remember is a click, not an archaeology dig.
Everything is findable
Every file in the agent's workspace and every past conversation is indexed as it's written — by keyword and by meaning, with results reranked so the best answer surfaces first.
The whole search stack runs alongside the platform. Nothing is shipped to a third-party service to make your history findable.
The part that changes how you work: agents search it too. Mine pull up decisions from weeks-old conversations in other rooms, unprompted, because looking it up is cheaper than guessing.
Lossless by default
Condensing is a view over the transcript, never a deletion of it.
Expand on demand
Any chapter reopens to the exact messages that produced it.
Self-hosted index
Keyword, semantic, and reranking all run on the same infrastructure.