HAM — Hierarchical Agent Memory

Fewer tokens. Greener AI.

HAM is a memory system for AI coding agents that reduces token consumption by up to 80%. Instead of loading your entire project context on every request, HAM scopes memory to the directory you're actually working in.

Less tokens. Faster agents. Lower costs. Greener AI.

Claude Code only — HAM is a skill for Claude Code and only works with Claude coding agents. It is not compatible with Cursor, Copilot, Windsurf, or other AI coding tools.

---

The Problem

Every time your AI agent starts a session, it re-reads everything. Your full project structure. Conventions it already learned. Decisions you already made. Context that has nothing to do with the current task.

A single bloated CLAUDE.md can eat 47% of your context window before the agent writes a single line of code.

That's wasted tokens. Wasted money. Wasted energy.

The Solution

HAM replaces one massive context file with small, scoped memory files at each directory level. Your agent reads only what it needs for the directory it's touching.

project-root/
├── CLAUDE.md                  # Global: stack, conventions (under 250 tokens)
├── src/
│   ├── CLAUDE.md              # Shared src patterns
│   ├── api/
│   │   └── CLAUDE.md          # API auth, rate limits, endpoint patterns
│   ├── components/
│   │   └── CLAUDE.md          # Component conventions, styling rules
│   └── db/
│       └── CLAUDE.md          # Schema context, query patterns
└── .memory/
    ├── decisions.md           # Architecture decisions with rationale
    └── patterns.md            # Implementation patterns

The agent reads 2-3 small files instead of one massive context dump. Your starting context drops from thousands of tokens to hundreds.

---

Before & After

| | Before HAM | After HAM | |---|---|---| | Context per prompt | 4,000 - 12,000 tokens | 800 - 2,400 tokens | | 50-prompt session | 200K - 600K tokens | 40K - 120K tokens | | Context window used at start | Up to 47% | Under 10% | | Token reduction | — | Up to 80% |

---

Why This Matters

For Your Wallet

Fewer tokens = lower API bills. Teams running agents at scale see the savings immediately.

For Your Speed

Smaller context = faster responses. Your agent spends less time processing irrelevant information and more time writing code.

For The Planet

AI inference accounts for over 80% of AI electricity consumption. Every token generated requires compute, energy, and cooling. Reducing token waste isn't just efficient — it's a sustainability decision.

Data centers are projected to consume 945 TWh of electricity by 2030 — more than Japan's total consumption. AI is the primary driver of this growth. — International Energy Agency

HAM makes your AI usage greener by eliminating the tokens that never needed to exist.

---

Installation

git clone https://github.com/kromahlusenii-ops/ham.git ~/.claude/skills/ham

This installs HAM globally — it's available in every project you open with Claude Code.

Add to a specific project only

If you prefer to scope HAM to a single project instead of installing it globally:

cd your-project
mkdir -p .claude/skills
git clone https://github.com/kromahlusenii-ops/ham.git .claude/skills/ham

Project-level skills live in .claude/skills/ and are only active in that project.

Quick Start

Open Claude Code in your project directory and say:

go ham

That's it. HAM auto-detects your stack, scans your project structure, and generates scoped CLAUDE.md files across your codebase. No manual setup required.

After setup, say HAM savings to see your token and cost reduction.

Updating

cd ~/.claude/skills/ham && git pull

---

How It Works

HAM follows three principles:

Scope, don't dump. Every piece of context lives in the most specific directory it applies to. Global conventions in root. API patterns in the API folder. Component rules in the components folder.

Read small, read relevant. The agent loads root context + the target directory's context. Two to three small files instead of the entire project.

Self-maintaining memory. Decision files and pattern logs update as the agent works. The root CLAUDE.md instructs the agent to read before coding and write before closing — context stays fresh without manual maintenance.

Commands

Setup

| Command | What it does | |---|---| | go ham | Set up HAM in your project (auto-detects everything) | | ham update | Update HAM to the latest version | | ham status | Show HAM version and setup status | | ham route | Add/update Context Routing in root CLAUDE.md |

Analytics

| Command | What it does | |---|---| | ham dashboard | Launch the web dashboard at localhost:7777 | | ham savings | Show token and cost savings report | | ham carbon | Show energy and CO2e efficiency stats | | ham insights | Generate insights and write actionable items to inbox |

Benchmarking

| Command | What it does | |---|---| | ham benchmark | Compare baseline vs HAM task performance | | ham baseline start | Begin 10-task baseline capture (no HAM memory loading) | | ham baseline stop | End baseline early, keep partial data | | ham metrics clear | Delete all benchmark data |

Maintenance

| Command | What it does | |---|---| | ham audit | Check memory system health | | ham commands | Show all available commands |

Dashboard

Say HAM dashboard (or HAM sandwich) to launch an interactive web dashboard at http://localhost:7777 that visualizes your actual Claude Code session data.

The dashboard shows:

• Token savings — estimated tokens and cost saved by HAM, comparing HAM-on vs HAM-off sessions
• Task benchmarking — baseline vs HAM performance comparison on the Overview tab
• Daily trends — charts of input tokens, cache reads, and cost over time
• Directory breakdown — which directories you work in most and their HAM adoption
• Session history — every session with model, duration, token counts, and HAM status
• Context health — which directories have CLAUDE.md files (green), which are stale (amber), and which are missing them (red)

Data is parsed directly from Claude Code's session JSONL files at ~/.claude/projects/ — no external services, no database.

Manual launch

If you want to run the dashboard outside of Claude Code:

# From your project directory
node ~/.claude/skills/ham/dashboard/launch.js [--port 8080]

The launcher auto-installs dependencies and builds the frontend on first run. Default port is 7777.

---

Git

HAM automatically adds its generated files to .gitignore during setup. All HAM files are local to your machine — they won't be pushed to your repo. ham remove cleans up the .gitignore entries.

---

HAM Pro

Scale to your whole team — HAM Pro adds multi-agent support for Cursor, Copilot, Windsurf, Gemini, and more.

Learn more →

---

License

MIT

---

Built by @kromahlusenii-ops

Saving tokens. Saving money. Saving energy.