How I built this

Ben's Bites published a walkthrough of setting up a new personal agent and a deliberately lightweight memory system. The author said around 700 readers reported that they either use or want to use a personal agent, and that roughly 50% of respondents to the previous day's poll wanted to know how it works in Claude. The author wrote that personal agents work in much the same way across tools, describing them as files, folders, tools and instructions. The stated motivation for rebuilding was that the existing personal agent was messy and often surfaced irrelevant material.

The author sketched a starting file set before beginning: AGENTS.md for main instructions and pointers; code.md for building preferences such as Vercel and Supabase; todos.md for current work; memory.md as a pointer file to other memory files covering the author, travel preferences, Ben's Bites and fund work; and log.md for a session log. A new folder named ~/bitess was created and a new thread started inside it.

The agent recommended merging building preferences into the main instruction file, which the author rejected because the personal agent is used for more than building. The author attributed the suggestion to using Codex, described as the coding variant of the agent, whose system instructions state it is a coding agent. The agent also recommended dropping log.md because git history records most work; the author accepted this after explaining that git saves versions of files as commits, that changes can be reviewed as diffs, and that a folder containing git is called a repository or repo.

On memory, the author said the previous folder instructed the agent to log important context, but this produced steered responses that kept the agent in a preferred lane rather than helping brainstorm new directions. The author is considering the smallest possible files and files' update when something changes.

For recalling chats, the author said agents store sessions in hidden dot files searchable by the agent. Codex mentioned SQLite, a database, as a way to search sessions without loading files into context; the author may try it later but for now will let the agent search threads directly. Files ended up with memory organised in a sub-folder, and CLAUDE.md contains only text directing Claude to read @AGENTS.md. The author distinguished instructions from skills, describing skill.md as a file of instructions and specific task bots as chat sessions with specific instructions and their own memories plus possible access to shared memory higher in the folder tree. The author plans to stick with one agent, small files and folders, and skill files for specific tasks.