Loading today’s stories
Product1 distinct publisher3 min readUpdated
The personal agent is a folder, not a model: four files and less memory than you thought
Ben's Bites rebuilt its author's personal agent from scratch. The lesson from version one was that automatic memory made the agent steer him instead of helping him.
The Product Desk · Product desk
What happened
- Around 700 Ben's Bites readers told the author they either use or want to use a personal agent.
- About 50% of respondents to the previous day's poll wanted to know how a personal agent works in Claude.
- The author says personal agents work pretty much the same way across Claude, ChatGPT and similar tools: it is just files, folders, tools and instructions.
- He says his own existing personal agent is pretty messy and often mentions stuff that is irrelevant, so he set up a fresh one.
- Before starting, he sketched five files for the new folder: AGENTS.md, code.md (building preferences such as Vercel and Supabase), todos.md (current work), memory.md, and log.md (a log of every session).
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
Retained memory failed here in a specific way, and it is worth being precise about it. The previous folder told the agent to write down anything that looked like important context about him or his work [13]. What came back, when he wanted to explore a new direction, was steering: the agent kept citing what the file said he liked and staying in that lane [14]. The instruction worked exactly as written. Written was the problem.
The same mechanism shows up inside a single session, at a shorter timescale. He floated auto-saving early on, then later asked the agent to review recent chats and describe what he actually uses it for. It came back still holding the auto-save idea and proposing another, more complicated auto-commit instruction [17]. Text in the context window outweighs the thing you asked for two messages ago, whether that text arrived from a memory file or from your own discarded suggestion.
The file layout is where the argument gets settled. Five files went into the sketch [5]. The agent contested two of them: merge the building preferences into the main instructions, and drop the session log because git history already records the work [9][11]. He took the second and refused the first, which leaves four [1][2]. His reason for refusing is the useful part. He is working in Codex, whose system instructions say it is a coding agent, so its advice about how to organise a workspace came out shaped like a coding workspace [10]. About half the readers who answered his poll wanted to know how this works in Claude specifically [2]; his answer is that the vendors work much the same way, being files, folders, tools and instructions [3]. On that reading, one line in AGENTS.md saying questions get answers, not changes [6] is doing more work than the choice of model behind it.
One loose end is left open, and it is the one an operator will hit first. The log file was dropped because git already stores versions of files, diffs and commits [12], and a duplicate is not worth the space [11]. But the request he says he makes most often is "what did we talk about last week re: [thing]" [18], and a commit history of file changes is not a record of conversation. He notes that agents save all chat sessions to files, and the published text breaks off mid-sentence there [19], with a dedicated memory post promised later [20].
What survives is a rule rather than a stack: pick the smallest files with the least context, know what is in them, and edit when something changes or when the agent starts saying things you do not like [16]. Memory, on this evidence, is not storage. It is a standing instruction about what gets re-read, and the first version of his agent read too much [4].
Loading claim ledger
Loading source directory links
Loading share composer
Loading topic controls
Loading related stories