Harness MaxxingDocs
All pages

Using Harness Maxxing

Memory

Claude has no memory between conversations – except what is written in files. Harness Maxxing organises that memory in plain markdown files that Claude Code loads automatically.

Three levels

/root/maxxing/
  CLAUDE.md  memory.md                         workspace (global)
  projects/acme/
    CLAUDE.md  memory.md                       project acme
    topics/blog/
      CLAUDE.md  memory.md                     topic acme / blog

Claude Code loads every CLAUDE.md from the session folder up to the workspace root, and each CLAUDE.md contains the line @memory.md, which imports its memory file. A session in projects/acme/topics/blog/ therefore reads all six files.

File Contains Who writes it
CLAUDE.md stable instructions: what the project/topic is, how to work on it. The first paragraph under the title is the description shown in the UI. you (or Claude when you ask)
memory.md what was learned: decisions, preferences, state of the work, useful paths and URLs Claude, through the memory skill

What goes where

  • Topic memory (most common): the state of this work stream, its decisions and conventions.
  • Project memory: what is true for every topic of the project.
  • Global memory: your preferences and rules valid everywhere (language, tone, "never push without asking"...). Keep it short – it is loaded in every session.

Not in memory: chit-chat, one-off questions, step-by-step logs, and never secrets (put them in Settings -> Keys and mention only the key name).

The memory skills

  • memory – at the end of a significant exchange Claude records durable facts in the right file: short dated bullets, replacing outdated facts instead of piling them up. You can also ask explicitly: "remember that ...".
  • memory-compaction – keeps files short and consistent. A scheduled job "Memory compaction" runs it every day at 05:00 (see Crons; it starts once the onboarding is finished with Claude signed in). It acts on files longer than the line limit (Settings -> Memory, 200 lines by default) or redundant, contradictory or outdated: it merges duplicates, summarizes old history and reorders each file (rules and preferences, current state, decisions, references, context). It only edits memory.md files.

Second brain cleaner (Settings -> Memory)

  • Maximum length per file: the line limit above (or No limit).
  • Clean now: starts the compaction immediately, in a session at the workspace root that you can watch.
  • The list shows every memory.md with its line count; files over the limit are in red.
  • From a shell: maxxing memory prints the same list.

Editing memory yourself

Memory files are ordinary files: open them in Files, edit, save. Changes apply to the next turn of every session in that scope.

Backup

Memory is the most valuable thing on the server. Enable the Backup to push it to a private GitHub repository every day, with history.

← SessionsFiles →