Skip to main content
skill CX Codex

memory — Claude Code skill: keeping agent memory clean

A skill that sets the discipline for a node's persistent memory: what to write, how to split pages, when to prune. Useful when your agent hoards history instead of knowledge.

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add plasma-ai/fractal --skill memory -a claude-code
Open on GitHub

What memory does

Node memory lives in $MEMORY_DIR and acts as the node's durable store. The skill covers reading it during orientation and folding findings back before an iteration ends. Sync can also write to memory when something important arrives over radio.

The core rule: write knowledge, not history. No iteration numbers, no timestamps — a reader should not be able to tell how many runs have happened. Before COMMIT the skill says to scan memory and any page you promote and rewrite those markers out, because promotion carries your habits onto shared surfaces. Organize by topic, not time: update the existing page for a topic rather than appending a new entry.

Your first write creates the layout — one page per topic (environment, decisions, state, todo, rename to fit the work), never a single grab-bag page. A page without desc: frontmatter stays invisible to wiki map, so after adding or moving pages run wiki update --path=$MEMORY_DIR. Keep exactly one present-tense state page, overwritten in place. In-wiki links are written prefix-free: [[state]], not [[./state]].

What it helps with

  • split node memory into topical pages instead of one grab-bag file
  • strip iteration numbers and timestamps out of memory
  • collapse a finished phase's pages into durable facts
  • make pages show up in wiki map with desc: frontmatter
  • keep a working todo list instead of a done-log

Example prompts for the agent

  • read $MEMORY_DIR and tell me where we are
  • fold this iteration's findings into memory and clean out the iteration numbers
  • run wiki update --path=$MEMORY_DIR so the new pages get frontmatter

How to install memory

1Skills CLI (Vercel)

npx skills add plasma-ai/fractal --skill memory -a claude-code

2Manually: put the files in place

~/.claude/skills/memory/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/memory/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about memory

How do I install memory in Claude?

Run npx skills add plasma-ai/fractal --skill memory -a claude-code. The skill file lands at ~/.claude/skills/memory/SKILL.md.

What does the memory skill do?

It describes how to maintain a node's persistent store in $MEMORY_DIR: what to write, how to organize by topic, and when to delete what is no longer true.

Why is my page missing from wiki map?

It probably lacks a desc: frontmatter key — a near-miss like description: does not count. After adding or moving pages, run wiki update --path=$MEMORY_DIR.

Source text

SKILL.md · Apache-2.0
---
name: memory
description: How to maintain node memory well -- the node's persistent knowledge store.
---

# Memory

Memory (`$MEMORY_DIR`) is the node's durable brain. Read it when you orient and
fold findings back before each iteration ends. Sync may also write to memory
when crucial information arrives via radio. See the `wiki` skill for how memory
relates to the shared project wiki; this doc is the discipline for keeping
memory useful.

Run `wiki --help` and `wiki <command> --help` for the CLI.

## Conventions

- **Write knowledge, not history.** Never reference iteration numbers,
  timestamps, or chronological markers -- a reader shouldn't be able to tell how
  many iterations have run. Before COMMIT, scan memory and any page you promote
  for iteration numbers and run labels and rewrite them out -- promotion carries
  your habits onto shared surfaces.
- **Organize by topic, not time.** Update the existing page for a topic; don't
  append a new entry.
- **Start with a topical layout.** Your FIRST memory write creates the layout --
  one page per topic (`environment`, `decisions`, `state`, `todo` -- rename to
  fit the work; stubs are fine), never a single grab-bag page; when a page
  starts covering two topics, split it. A leaf expecting to finish within an
  iteration or two may keep just the `state` page.
- **One current-state page.** Keep exactly one present-tense `state` page for
  where-am-I / what's-next, overwritten in place -- no "prior run" sections;
  per-iteration narrative belongs in the plan post-mortem, never memory.
- **Pages need `desc:` frontmatter.** A page without it (a near-miss key like
  `description:` counts as without) is invisible to `wiki map`; after adding or
  moving pages, run `wiki update --path=$MEMORY_DIR` so they gain frontmatter
  and index entries. Quote a one-line `desc:` that contains `: ` or ` #`.
- **Fold at phase ends.** When a phase or a child's run ends, collapse its pages
  into durable facts and delete the rest -- memory carries what is still true,
  not what happened.
- **No append-only logs.** If you're adding dated entries, stop -- replace
  outdated content with current understanding.
- **Todo lists are living state.** Keep your private working checklist here as
  current open items, pruned as they complete -- never a done-log. A todo list
  other nodes should see and track belongs in the project wiki instead.
- **Keep indexes lean.** Keep each `_index.md` under ~100 lines below the `***`;
  factor overflow into child pages.
- **Wikilinks stay within one wiki.** Reference anything outside this wiki --
  the project wiki, source files, configs -- in plain text or backticks, never
  as a wikilink; `wiki lint` notes a wikilink that points outside the wiki.
  Write in-wiki links prefix-free (`[[state]]`, never `[[./state]]` or
  `[[../state]]`).

© plasma-ai · Apache-2.0 · plasma-ai/fractal

Similar skills

All skills for Agent setup: memory, context, orchestration →

Project sponsors

Companies that keep this analytics open
Gold sponsor

Impressions, clicks and code copies — open statistics. We keep the numbers public for the same reason NeBlask itself is public. Become a sponsor