axme-code

AXME Code — Claude Code Plugin

Alpha — This plugin is in active development. Expect breaking changes between versions. Report issues.

Persistent memory, architectural decisions, and safety guardrails for Claude Code. Your agent starts every session with full project context — stack, decisions, patterns, safety rules, and a handoff from the previous session.

Main Repository · Website · Issues

---

What It Does

Claude Code forgets your project every session. AXME Code fixes that.

• Persistent Knowledge Base — oracle (stack, structure, patterns, glossary), architectural decisions with enforcement levels, accumulated memories from past sessions
• Safety Guardrails — hooks intercept dangerous commands before execution (not prompts — hard enforcement at the harness level). Blocks git push --force, rm -rf /, npm publish, direct commits to main, and more
• Session Continuity — handoff system tells the next session exactly where work stopped, what's in progress, and what to do next
• Automatic Learning — background auditor extracts memories, decisions, and safety rules from session transcripts when you close the window
• 19 MCP Tools — axme_context, axme_save_decision, axme_save_memory, axme_safety, axme_backlog, and more

You keep using Claude Code exactly as before. AXME Code works transparently in the background.

Install

From terminal:

claude plugin install axme-code

From Claude Code interactive CLI:

/plugin install axme-code

First Use

After installing the plugin, open any project in Claude Code. On the first session:

1. The plugin creates a CLAUDE.md section that instructs the agent to call axme_context
2. The agent will say the project is not initialized and run axme-code setup --plugin
3. Setup scans your project (uses an LLM to analyze your stack, extract decisions, detect safety patterns) and creates the .axme-code/ knowledge base
4. From the next session onward, the agent automatically loads the full knowledge base

Setup requires a Claude subscription (the LLM scan uses Claude Sonnet). If the scan fails, a deterministic fallback creates a minimal knowledge base from file system inspection.

How It Works

Hooks (automatic, transparent)

Hook	When	What
SessionStart	Session opens	Installs SDK dependency, creates CLAUDE.md if missing
PreToolUse	Before every tool call	Checks safety rules, blocks violations
PostToolUse	After Edit/Write	Tracks which files changed
SessionEnd	Session closes	Spawns background auditor to extract knowledge

MCP Server

Provides 19 tools for reading and writing the knowledge base. The agent calls these tools naturally during work — no manual intervention needed.

Storage

All data lives in .axme-code/ in your project root (gitignored automatically):

.axme-code/
  oracle/           # stack.md, structure.md, patterns.md, glossary.md
  decisions/        # Architectural decisions with enforce levels
  memory/           # Learned mistakes (feedback/) and validated patterns (patterns/)
  safety/rules.yaml # git + bash + filesystem guardrails
  backlog/          # Persistent cross-session task tracking
  sessions/         # Session metadata
  plans/            # Handoff files for session continuity
  worklog.jsonl     # Structured event log

Human-readable markdown and YAML. No database, no external dependencies.

Requirements

• Claude Code (CLI or VS Code extension)
• Node.js >= 20
• Claude subscription (for LLM-powered setup and session auditor)

Alpha Status

This is alpha software (v0.x.y). What that means:

• Breaking changes may happen in minor version bumps (0.x.0 -> 0.x+1.0)
• Storage format may change between versions (.axme-code/ is local and gitignored, so upgrades won't affect your repo)
• Bug reports are welcome — open an issue
• Feedback — we're actively iterating based on user feedback. Tell us what works and what doesn't

Auto-Sync

This repo is auto-synced from the main axme-code repository on each release. Do not edit files here directly — changes will be overwritten.

License

MIT

---

Website · Main Repo · Issues · contact@axme.ai