Claude Memory Engine

Not just memory — it learns.
Learn from mistakes. Learn to improve.
AI can be a student too, growing through every cycle.

Built with hooks and markdown. No database. No external API.
Just scripts and files. Nothing hiding.

English  |  繁體中文  |  日本語

WHAT — Every new conversation, Claude starts from zero

• That bug you spent 30 minutes on last session — it hits the same wall again
• Your preferences, your project rules — gone the moment a new session starts
• Switch from Project A to Project B — it can’t tell which is which
• Long conversations get fuzzy — important decisions vanish after compression
• Memory files pile up — no one organizes them, they just keep growing
• Your computer dies — local memory gone, no backup

Memory tools can help it “remember.” But remembering is not the same as learning.

WHY — Because it learns

Memory Engine doesn’t just help Claude remember — it teaches Claude to learn like a student:

• Mistakes don’t repeat — it saves both the problem and the fix
• Switching projects doesn’t mean starting over — it knows what you’re working on
• It gets better over time — each cycle, it understands you a little more
• You can see how it learns — everything is markdown and JS, no black box

HOW — Through the Student Loop

• Student Loop — 8-step learning cycle, like cramming for finals but it keeps getting better
• Smart Context — auto-loads the right project’s memory based on your working directory
• Auto Learn — saves both the problem and the fix when it hits a wall, won’t repeat the same mistake

:brain: The Student Loop

Think of it like exam prep. I’m trying to make Claude Code act like a student cramming for finals — take notes after every class, organize them, review for patterns, build an error notebook, and do a big end-of-term review. Each cycle, it gets a little better.

In class (automatic, runs every session)

There is no real “end” to a Claude Code conversation — it might close, idle out, or get compressed. So Memory Engine doesn’t rely on any single moment. Instead, it saves at three different points:

1. Every 20 messages (mid-session-checkpoint) — saves a checkpoint + mini analysis. The most reliable save point, because it counts messages itself
2. Before context compression (pre-compact) — fires right before context is compressed. Saves a snapshot, detects pitfalls, runs backup. This is when context is fullest, so pitfall detection is most accurate here
3. When the conversation ends (session-end) — saves a final summary + backup. Nice to have, but not guaranteed to fire (window might just close)

You don’t need to remember to run any command before closing — your important stuff is already saved before you close.

On top of that, Claude also:

• Takes notes — records what was done, which files changed, key decisions made
• Links them — tags the project, connects to previous notes
• Spots patterns — scans for pitfall signals (retrying 5+ times, errors followed by fixes, user corrections)

Final exam review (manual, run /reflect)

After a few days of notes, run /reflect and Claude will:

4. Review — read the past 7 days of notes and pitfall records, mark what’s still useful and what’s outdated
5. Refine — apply four decision questions: Keep it? -> Condense it? -> Already covered by a rule? -> Delete only as last resort
6. Re-study — re-analyze the cleaned-up data to find patterns that were buried in noise
7. Slim down — list items that can be removed, wait for your confirmation before deleting anything
8. Wrap up — produce a report: what was learned, what changed, what to watch for next cycle

This isn’t a one-time thing. Each cycle makes the notes sharper, the patterns clearer, the mistakes fewer. It’s a loop that keeps improving.

:pencil2: Correction Cycle

Some mistakes don’t show up in error logs. You correct its output, and only then does it realize — “oh, that was wrong.” These mistakes don’t get remembered automatically. Unless someone builds it an error notebook.

Record (/analyze, manual — run right after you correct something)

• You fix its work, type /analyze
• It compares both versions against existing rules
• Known rules it missed → logged, counted
• Patterns not in the rules yet → distilled into new ones
• The sooner you run it, the fresher the context

Review (automatic: before each task / manual: type /correct anytime)

• Before starting work, it scans the error notebook automatically
• Not re-learning — just a reminder: “I got this wrong last time, don’t repeat it”
• Want to review on your own? Type /correct — no need to wait for a task or a cycle

Clean up (/reflect step 6, manual)

• Periodically scan the full notebook
• Same mistake 3+ times → upgrade to a hard rule
• Already internalized → mark cleared, free up space

But you know that from here on, your AI has grown a little more.

:detective: Smart Context + Auto Learn

Smart Context (automatic, no config needed)

• Whatever folder you’re working in, it loads that project’s memory
• Switch projects and it switches automatically

Auto Learn (automatic, on session end)

• Hit a wall and figured it out? It saves both the problem and the fix
• Reminds itself next time a new conversation starts
• Same kind of mistake 3+ times across different days → suggests writing it into permanent rules

:link: Day-to-day tools

Memory and learning are the core, but day-to-day work needs more:

:arrows_counterclockwise: Cross-device Sync (rewritten in v2.0)

Memory Engine supports cross-device sync through a GitHub repo. Set it up once, and your memory works on every machine.

How it works:

1. /backup pushes local memory to your private GitHub repo (memory-backup.sh push)
2. /sync does bidirectional sync — pull remote updates, then push local changes (memory-backup.sh sync)
3. /recover on a new device pulls everything back and distributes to all local projects (memory-backup.sh pull)

What this means: Switch laptops, reinstall your OS, set up a new workstation — run /recover and Claude picks up right where you left off. No re-explaining your preferences, no lost context.

The GitHub repo is private by default. Your memory never touches any external service beyond your own GitHub account.

v2.0 — Full Environment Engine

v1.6 only synced memory files, which meant your second machine forever ran an older CLAUDE.md, missing slash commands, and stale hooks. v2.0 rewrites the sync as a full environment engine — the entire ~/.claude/ working environment travels with you:

Core architecture: SYNC_TABLE

A single source-of-truth table inside memory-backup.sh. Both push and pull read the same table, fixing the asymmetric bug where push synced X but pull silently skipped X. Adding a new sync target is one line:

SYNC_TABLE=(
  "$CLAUDE_DIR/CLAUDE.md|$REPO_DIR/CLAUDE.md|file|CLAUDE.md (top-level rules)"
  "$CLAUDE_DIR/commands|$REPO_DIR/commands|dir-md|slash commands"
  # ... add a line, push/pull both pick it up
)

v2.0 Features

:package: Installation

Step 1 — Create a GitHub repo for memory backup (cross-device sync):

Without a backup repo, /backup, /sync, and /recover won’t work. Memory only lives locally — if your machine dies, it’s all gone. With a repo, your memory works across devices.

gh repo create claude-memory --private
git clone https://github.com/YOUR_USERNAME/claude-memory.git ~/.claude/claude-memory

Step 2 — Copy files:

cp hooks/*.js ~/.claude/scripts/hooks/
cp hooks/*.sh ~/.claude/scripts/hooks/
cp commands/*.md ~/.claude/commands/
cp -r skill/ ~/.claude/skills/learned/memory-engine/

Step 3 — Create directories:

mkdir -p ~/.claude/sessions/diary
mkdir -p ~/.claude/scripts/hooks

Step 4 — Add hooks config to ~/.claude/settings.json:

{
  "hooks": {
    "SessionStart": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "command",
            "command": "node ~/.claude/scripts/hooks/session-start.js"
          }
        ]
      }
    ],
    "SessionEnd": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "command",
            "command": "node ~/.claude/scripts/hooks/session-end.js"
          }
        ]
      }
    ],
    "UserPromptSubmit": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "command",
            "command": "node ~/.claude/scripts/hooks/memory-sync.js"
          },
          {
            "type": "command",
            "command": "node ~/.claude/scripts/hooks/mid-session-checkpoint.js"
          }
        ]
      }
    ],
    "PreCompact": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "command",
            "command": "node ~/.claude/scripts/hooks/pre-compact.js"
          }
        ]
      }
    ],
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "node ~/.claude/scripts/hooks/pre-push-check.js"
          }
        ]
      },
      {
        "matcher": "Write",
        "hooks": [
          {
            "type": "command",
            "command": "node ~/.claude/scripts/hooks/write-guard.js"
          }
        ]
      }
    ]
  }
}

Step 5 — Restart Claude Code. Done!

:rocket: Quick Start

Done installing? Here’s what to do next.

1. Just start working — open Claude Code and go. session-start loads your last session’s context automatically
2. Just close when done — session-end saves a summary if it fires; mid-session-checkpoint and pre-compact already have your back
3. Want to remember something? — /save stores it in long-term memory
4. Switching windows? — /handoff passes your progress to the next window
5. After a few days — /reflect reviews your notes, finds patterns, cleans up

That’s it. Everything else runs in the background.

:zap: Token Impact

Memory Engine adds almost no token overhead to your daily usage.

SKILL.md (136 lines) is a learned skill — Claude Code only loads it when relevant, not every conversation.

Bottom line: ~200–500 extra tokens at the start of each conversation. Everything else is zero unless triggered.

:wrench: Customization

:bulb: Design Philosophy

Why not a database?

• Markdown — you can open it, read it, edit it, git commit it
• Claude Code already reads .md natively — why add complexity?

Why not a Plugin?

• Plugins are black boxes
• Hooks + Commands are transparent — every .js file is right there to inspect
• Don’t like something? Change it. Think it’s unnecessary? Delete it
• Tools should be something you control, not something that controls you

:pray: Credits

All code was written from scratch. No code was copied, forked, or adapted from any source project.

claude-memory-engine/
  hooks/
    session-start.js          # New session -> load recall + smart-context + handoff
    session-end.js            # Session end -> save summary + backup (best-effort)
    pre-compact.js            # Context compression -> snapshot + pitfall detection + backup
    shared-utils.js           # Shared functions (transcript, pitfall, backup)
    memory-sync.js            # Every message -> cross-session memory sync + handoff
    write-guard.js            # Before file write -> sensitive file warning
    pre-push-check.js         # Before git push -> safety check
    mid-session-checkpoint.js # Every 20 messages -> checkpoint
    memory-backup.sh          # Bidirectional sync (push/pull/sync) — v1.6
  commands/
    save.md / 存記憶.md        # Save memory across sessions
    reload.md / 讀取.md        # Load memory
    todo.md / 待辦.md          # Cross-project tasks
    backup.md / 備份.md        # Push to GitHub
    sync.md / 同步.md          # Bidirectional sync
    diary.md / 回顧.md         # Reflection diary
    reflect.md / 反思.md       # Pattern analysis
    learn.md / 學習.md         # Pitfall learning
    check.md / 健檢.md         # Quick health check
    full-check.md / 大健檢.md   # Full audit
    memory-health.md / 記憶健檢.md
    memory-search.md / 搜尋記憶.md
    recover.md / 想起來.md
    compact-guide.md / 壓縮建議.md
    handoff.md / 交接.md        # Session handoff
  skill/
    SKILL.md
    references/
      smart-context.md
      auto-learn.md

Requirements

• Claude Code (with hooks support)
• Node.js 18+
• Zero dependencies

License

MIT — see LICENSE for details.

Made by HelloRuru