Every conversation with a raw LLM starts from zero — it doesn’t know you, your projects, or what you decided yesterday. The fix isn’t exotic: files. Plain markdown, loaded into context at session start, updated as things change. Our assistant runs on this exact pattern, and so do the major agent harnesses.

The four-file architecture

  1. soul.md — who the assistant is: personality, tone, standing rules. Written once, edited rarely. Keep it under ~5,000 characters — this text rides along with every request, so bloat here taxes every response.
  2. user.md — who you are: preferences, context, how you like to work, what annoys you. (“Prefers direct answers”; “always metric”; “never reformat my code unasked.”)
  3. memory.md — the living file: current projects, recent decisions, open questions. This one changes constantly.
  4. Skills (optional folder) — repeatable how-to files for tasks it does often.

The assistant reads all of these when a session starts. That’s the entire trick — the “memory” is just context it re-reads.

The habits that make it actually work

  • Update memory at the end of anything significant. The command is literally “update your memory file with what we decided.” Assistants that write their own memory (with your review) stay current; ones that rely on you journaling for them don’t.
  • One fact per entry, dated when it matters. “Prefers Suno for scores (2026-07)” ages better than a paragraph of vibes.
  • Prune monthly. Stale memory is worse than no memory — the assistant will confidently act on your old preferences. Read the file sometimes; delete what’s no longer true.
  • Memory ≠ logs. Don’t append transcripts; distill decisions. “We chose X because Y” is memory; the 40-message debate isn’t.

Why this beats fancy solutions for most people

Vector databases and RAG have their place (searching thousands of documents), but for personal continuity, files win: you can read them, edit them, version-control them, and carry them between tools. When we migrated our assistant across agent frameworks — twice — the memory files moved with a copy-paste and the personality survived intact. Try that with an embedded vector store.

Level up

Once memory works, the same files become the foundation for multiple personas (per-character soul files), scheduled routines (the assistant reads memory, does the morning briefing), and real autonomy — an assistant that remembers is one you can delegate to. Ours runs a daily video on exactly this foundation.