I Give My Hermes Agents a Haircut Every Week

I run seven AI agents for my company. Each one keeps memory files so it can carry rules and context from one job to the next. That is useful until it is not. Hermes agents can learn from a job, write the lesson into memory, and keep moving. But those files only grow unless somebody goes back through them. I had never really audited what they were writing down. Claude found about 31,000 characters of notes loading into the fleet. Some were useful. A lot were duplicates. A few were flat-out wrong. One agent had overlapping notes about the same customers. Another had memory entries teaching a workaround I had banned a month earlier. One still had an old model lineup in its instructions. Nothing was broken because of one dramatic failure. The agents were just slowly carrying more clutter and bad assumptions into every new job. Now Claude runs a weekly audit. It checks questionable notes against the live system, merges duplicates, preserves legitimate exceptions, backs up the originals, and tests the cleaned memory before closing the job. The first serious pass took the fleet from 27,716 characters to 15,170. More importantly, the agents are not being taught old rules by their own memory files. One catch: Claude initially wanted to remove a note that contradicted the main instructions. It turned out to be an exception I had personally approved. So the audit now has to check who wrote a disputed note and why before it removes anything. I wrote up the full workflow here: mylifeinthestack.substack.com/p/dispatch-34-i-give-my-hermes-agents Curious how other people handle long-term agent memory once they have more than one agent running.