How I built this
Ben's Bites
Ben Bites built a fresh personal AI agent with tiny files and almost no memory. He says agents mostly need clear instructions — not a giant pile of context.
Based on reporting by Ben's Bites — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Ben Bites is rebuilding his personal agent from scratch, and this time he’s trying to keep it small on purpose. The setup starts with a new folder and a few plain-text files: AGENTS.md for the main instructions, code.md for building preferences, todos.md for current work, memory.md as a pointer to other memory files, and log.md for session history.
Then he started stripping things back. His agent suggested folding the coding preferences into the main instruction file and dropping log.md because git history already tracks changes. That made sense to him. If the folder is under git, the old versions are already there, which makes a separate log file feel redundant.
Memory is where he gets uneasy. He had been thinking about automatic memory, with the agent learning from daily and weekly use, but he says that often leads to responses that are too steered by old preferences. He’d rather use the smallest possible set of files, know exactly what’s in them, and update them when needed instead of letting the system keep accumulating context he doesn’t fully control.
He also pushed back on another suggestion: keeping a database of old chats. The agent can already search past sessions stored on the computer, often in hidden dot folders like ~/.agents/, ~/.codex/, or ~/.claude/. So instead of building extra machinery right away, he wants the agent to search the threads itself and answer questions like “what did we talk about last week?”
The bigger conclusion is less about one file and more about how these systems behave. Ben says agents are agreeable, which is a polite way of saying they’ll happily talk you into complexity. His current plan is one main agent, small files, and separate skills for repeatable tasks, while he figures out what memory should actually do.
My take — AI-written commentary, not fact-checked reporting
The cleanest agent setup is usually the least impressive one. People keep bolting on memory, logs, databases, and clever little sidecars because it feels serious, then act surprised when the thing turns into a junk drawer with a chatbot attached. The boring version — fewer files, clearer rules, more human judgment — is the one that tends to work.
Read more about this at: Ben's Bites