Agent memory is being rebuilt as plain documentation
'Agents don't need memory, they need documentation' (49945933) argues that structured docs beat bolt-on memory systems. Commenters brought real scars: one said a home memory tracker burned tokens on reading and updating, and kept stale info alive. Another cited Peter Naur's point that docs can't fully capture the mental model behind a program. Someone else described mapping every statement into a proposition tree.
Meta's Muse (49946526) shows the other side of the same shift: a hosted agent with root access to its own VM, a tailnet connection and a torrent seedbox. One commenter asked why that's different from what ChatGPT agents already do. The local-model thread (48542100) shows the cost side. People reported Qwen 3.6 27B performing about like Claude Haiku 4.5, while others on Apple M4 hardware found token speeds too slow.
The pattern: the agent stack is converging on boring things. Files, docs and a sandbox beat clever memory architectures, and cost per token decides where the work runs.
So what?
If you build agent products, invest in how your tool writes and reads durable documents before you invest in fancy memory. Local open-weight models are close enough for routine coding that pricing power at the big labs is under real pressure.