Context accumulation is not control
Buzz accumulates context extremely well and governs it not at all. There is no quality scoring, no freshness indicator, no pruning, no signal that one statement supersedes another. A decision reversed in week three sits in the channel next to the decision it reversed, carrying equal weight.
Everything an agent reads looks equally current and equally true.
Why it matters to you
This is the gap between Buzz and what a mature context system would do, and it is the specific reason a long-running workspace degrades. Early on, everything in the channel is fresh and the accumulation is pure benefit. Over months, the proportion of stale, superseded and abandoned material grows — and nothing marks it as such.
For projects measured in weeks this is manageable. For a workspace running for a year, it is the main thing that will quietly reduce quality.
How to apply it
State supersession explicitly. "We are no longer doing X; the decision is now Y" is a sentence the workspace needs and will not infer.
Keep the canvas current rather than append-only — it is the one place where you can simply remove what is no longer true. See Maintain the channel canvas.
Prune on a rhythm. End of a project phase is a natural point: what in this channel is now misleading?
Date anything that will age — data feeds especially, where an agent will reason confidently from figures that stopped arriving weeks ago.
Prefer new channels to reused ones when a project genuinely restarts. A clean room is cheaper than curating a dirty one.
If you ignore this
Left alone, context rot produces confident wrong answers rather than obvious failures. An agent citing a superseded decision sounds exactly like an agent citing a current one, and the error surfaces downstream — in a deliverable, or in front of a stakeholder.
The cost of prevention is minutes per phase. The cost of the failure is credibility.
Examples
A filing cabinet nobody ever weeds. Everything is in there, including the version you replaced, and nothing on the folder says which is which.
A project's scope is widened in week two. In week six an agent produces analysis against the original scope, because both statements are in the channel and neither is marked as superseded.
Assuming that because the workspace has "all the context", its answers improve indefinitely. Past a point, more context without curation makes them worse.
Verified against Buzz v0.5.20 · Updated 2 Sep 2026