Claude Code Digest — 2026-03-26 20:13:29
What the docs reveal
Anthropic imposed hard limits on automated agent memory. Claude writes session learnings to MEMORY.md, but now loads only the first 200 lines or 25KB upon startup. It truncates all remaining content. In contrast, Claude continues to load CLAUDE.md entirely.
Unbounded memory bloats the context window. When an agent appends learnings to a single file over dozens of sessions, token requirements explode. This bloat degrades reasoning, increases latency, and wastes money. By capping MEMORY.md, Anthropic forces developers to groom context. The engineering team prioritized immediate reasoning capability over infinite historical recall.
This limit establishes a strict architectural hierarchy. CLAUDE.md functions as the project's foundational constitution. MEMORY.md operates as a volatile scratchpad. Anthropic trusts developers to maintain concise rules in CLAUDE.md, but they block the agent from generating unregulated memory logs.
You must adjust your documentation workflow to handle these constraints. If you expect Claude to reference past architectural decisions automatically, audit your MEMORY.md file size. Claude leaves all data past the 200th line in the dark.
Refactor your project context to survive truncation. Extract deep technical details from MEMORY.md and isolate them in dedicated topic files. Claude can find and read separate markdown files when a task requires specific knowledge. Reserve CLAUDE.md for critical syntax rules, standard terminal commands, and core project boundaries. Treat MEMORY.md as temporary cache, not persistent storage.