A Memory Hole That Blew Up at Dawn

While running OC on a Mac mini, after turning on 4.5's dreaming, I found in the morning that DREAMS.md was connecting project decisions made 3 days apart.

Previously, I had to manually select what went into long-term memory, which took a lot of effort, but dreaming started taking over part of that role.

However, 3 weaknesses also surfaced during operation.

• MEMORY.md getting overwritten during the day

  • There was an issue where the agent would clean up overnight-promoted entries the next day using the edit tool, deleting 2 of them in the process.
  • As a solution, chmod 444 was applied to core bootstrap files, and an append-only rule was applied to memory writes.
  • MEMORY.md was left at 644 so dreaming could keep writing to it, while the agent's edit permissions were restricted.

• No way to verify whether overnight work actually completed properly

  • Since dreaming runs at 3am and the user wakes up at 7am, it's hard to immediately know whether all 3 stages actually finished.
  • DREAMS.md had to be opened directly to check, and to address this, a cron check that runs before dreaming was added.
  • It checks file size, configuration, and whether the daily log exists, catching problems in advance that would otherwise only be discovered days later.

• The quality of daytime logs determines dreaming quality

  • When editing daily logs with the edit tool, exact string match would break, corrupting entries.
  • Switching this to append-only write made daily notes much more stable, and improved the material available for dreaming's light phase.
  • On top of this, proactive /compact was set at 60% to avoid emergency compaction, and memoryFlush secured time to save before compaction.

Additionally, by fixing the morning boot sequence and creating a roughly 500-token summary card loaded into the system prompt instead of re-reading all files, token usage dropped significantly from about 262k to under 12k.

This structure has been organized at github.com/aristotle-agent/aristotle, though it hasn't been published to npm yet. Lastly, the biggest remaining question is how to protect MEMORY.md more cleanly.