The myth: the more your AI agent remembers, the smarter it gets. Give it a memory file, let it write down everything it learns, and every run starts a little wiser than the last.
I believed this. My agents’ rulebook literally says “always read MEMORY.md at the start of every session.” So this morning I measured what that file has turned into after five months of my agents writing to it every day.
The receipt
- MEMORY.md is 9,888 lines and 2.3 MB. At the rough rule of 4 characters per token, that’s about 569,000 tokens of “memory.”
- The file’s own title sits on line 3,610. Everything above it is 330 run logs, newest first, stacked on top of the actual memory.
- The file reader my agents use returns 2,000 lines by default. That’s about 86,000 tokens, or roughly 15% of the file. It covers October 1 to today and nothing else.
- A default read never reaches the sections that matter most. Site & Infrastructure starts at line 4,657. Audience Insights is at 4,695. Technical Notes is at 4,758.
- 114 notes dated October or later sit below line 2,000, some of them written at the very bottom of the file. They were written last month and haven’t been read since.
So the agent “remembers” 569k tokens and reads the top 15%, which is mostly last month’s diary. The hard-won lessons from June and July are in there. They just aren’t anywhere the agent looks.
Why more memory makes agents worse, not smarter
Memory only helps at the moment it reaches the model. Anything written but never loaded is just storage. And a growing file causes three quiet problems:
⚡ GET THE AI EDGE
Weekly AI tips that actually save you time and money. No fluff, no hype — just what works.
- The good stuff sinks. When every run adds its log to the top, durable rules get pushed down by daily noise. The newest note wins, not the most important one.
- Old truths stop being true. A note from August about a dead API token is still in there, written with full confidence. An agent that does read it has no way to know it was fixed later that month. (I wrote about this in my instruction drift check: 7 claims in the rulebook had stopped being true.)
- Every token loaded costs room for the actual work. I measured the fixed overhead of one long-post run at about 19,000 tokens before it does anything (full numbers in MCP vs CLI: what actually reaches the model). An 86k-token memory read on top of that is context spent on yesterday’s logs.
Here’s the honest version of how this post got written. To avoid repeating a myth I’d already busted, my agent didn’t read its memory. It searched it for one phrase, “myths burned,” and found the list in a second. Search did the job that “read everything first” was supposed to do.
What actually works: small memory, searchable logs
- Split memory into two files. A short core file with facts that are true on every run: IDs, rules, known gotchas. Keep it under a few hundred lines and load it every time. Logs go in a separate file that is never loaded by default.
- Write rules, not diaries. When a run learns something, it should edit the rule in the core file (“Asana: create the task, then add the section in a second call”), not add a dated story about the day it found out.
- Give the core file a size cap. When it goes over, compact it: merge duplicates and delete anything no longer true. A memory you never prune turns into a pile of old notes.
- Search the rest on demand. At this size, a plain text search beats anything fancier. That’s the same conclusion I reached in my RAG pipeline guide: grep first, vector database only if grep can’t find it.
I haven’t restructured my file yet. Changing what every agent in the fleet reads at startup is a call I want to make myself, so my agent filed the numbers and the proposed split for my review. The original idea from my first memory tip still holds: memory is just a text file. The lesson is that someone has to edit that file, not just keep adding to it.
The takeaway: an agent doesn’t get smarter by remembering more. It gets smarter when the right few lines reach the model at the right moment. Check your agent’s memory file today: how long is it, and where do the important rules sit?
Want help designing agent memory that stays useful past month three? Book a strategy session.
📥 FREE: THE AI PLAYBOOK
The exact tools and workflows I use to run a one-person agency. 25 years of marketing experience distilled into an actionable guide. Yours free.
AI & Digital Marketing Strategist
With over 25 years in digital marketing, e-commerce, sales, and international business, Jon Jones is a seasoned entrepreneur dedicated to empowering small businesses through innovative AI solutions. His journey began in investment banking on Wall Street, guiding companies through capital acquisition and complex markets, and has expanded to include expertise in web development, brand design, product development, and courses and training.
Fluent in Mandarin and with a rich foundation in both finance and tech, Jon launched JonJones.ai to provide accessible, transformative tools for small businesses to streamline operations, generate leads, and build sustainable growth. His mastery of AI, automation, and digital strategy has enabled numerous businesses to bridge the gap between traditional practices and modern technology, ensuring they thrive in today’s digital landscape.
