Does a line in auto memory steer the next session like CLAUDE.md does? 3 of 4 runs vs 4 of 4, and 0 of 4 with neither

On a single line stored in Claude Code's auto memory, the model was successful in 3 out of 4 fresh sessions. In contrast, the same line in CLAUDE.md led to a first-attempt success in 4 out of 4 sessions. However, when neither memory system was utilized, none of the 4 sessions recovered the information. This experiment aimed to measure whether capturing a memory truly assisted later sessions, revealing a surprising result.

Claude Code stores auto memory in a project-specific directory, ~/.claude/projects/project/memory/, derived from the project's git repository path. The first 200 lines of the MEMORY.md file, or the first 25KB, are loaded at the start of every conversation. Topic files, such as user_role.md or feedback_testing.md, are not loaded at startup but read on demand when needed.

The experiment involved a project outside any git repository with a hidden environment variable check against a hash. A throwaway project was created, and a script was used to determine whether the PROJECT_TOKEN was valid, missing, or invalid. The measurement was based on the first line in the .attempts.log file, indicating whether the token was valid on the first attempt, after a failure, or never appearing.

The results showed that a line in the project's CLAUDE.md or MEMORY.md index had different effects on Claude Code's performance, with the MEMORY.md line only being successful in the first attempt and not in subsequent sessions.