agent-context-pack

Persistent memory for AI agents, version-controlled and synced through GitHub.

AI agents are stateless. Every new session, every new machine, every new agent starts from zero — no memory of what the project is, what decisions were made, or how things work. This skill solves that by giving every repo a structured context pack that acts as durable, shared memory any agent can read instantly.

Because it lives in git, this memory is:

• Persistent — survives across sessions, machines, and agent restarts
• Versioned — every change is tracked, reviewable, and reversible via PR
• Portable — any agent on any machine clones the repo and has full context immediately
• Collaborative — multiple agents and humans contribute to the same shared understanding
• Authoritative — one source of truth, not scattered across chat logs and local files

The Problem

Without persistent context, every agent session is groundhog day:

• Agent doesn't know what the repo is for or what files matter
• Agent doesn't know what decisions were made or why
• Agent doesn't know what it's allowed to do
• Agent doesn't know domain-specific terminology
• Agent on Machine B has no idea what Agent on Machine A learned yesterday
• Onboarding a new agent (or a new model) means re-explaining everything from scratch

Chat history is ephemeral. Local memory files don't sync. The knowledge dies when the session ends.

The Solution

Add two things to every repo:

1. CLAUDE.md at the root — the agent entry point
2. .context/ directory — structured context pack: project metadata, permissions, onboarding, architecture, glossary, decisions, and workflows

Any agent that lands in the repo reads CLAUDE.md, follows the reading list into .context/, and within 30 seconds has full project understanding — who owns it, how it's built, what the terms mean, what decisions were made, and what it's allowed to do.

Because it's in git:

• Agents on different machines read the same context
• Context updates go through PRs (reviewed, approved, merged)
• Architecture decisions accumulate as ADRs over time
• The glossary grows as the project grows
• New agents or models inherit everything previous agents learned

Your agents finally have a shared brain.

Use as a Template

Click "Use this template" on GitHub to scaffold a new repo with the context pack structure already in place.

Install as a Skill

Claude Code

gh repo clone JasonCZMeng/agent-context-pack
cp -r agent-context-pack/skill/workspace-context-pack ~/.claude/skills/workspace-context-pack

Claude Code auto-discovers it. When you set up a new repo or add context to an existing one, it uses this skill automatically.

Other Agents

Point your agent at skill/workspace-context-pack/SKILL.md as a reference document. The templates in context-pack-template.md are copy-paste scaffolds.

Quick Start (without installing)

• SKILL.md — the full process, standards, and templates
• context-pack-template.md — copy-paste scaffold for every file

What's in the Context Pack?

your-repo/
  CLAUDE.md                           # Agent entry point — ordered reading list
  .context/
    project.yaml                      # Name, type, tech stack, key files
    permissions.yaml                  # What AI can/can't do
    context/
      README.md                       # Index of context files
      onboarding.md                   # What this repo is, first steps
      architecture.md                 # Repo structure, how pieces fit
      glossary.md                     # Domain-specific terms (grows over time)
      decisions/
        ADR-0001-template.md          # Architecture Decision Record template
      workflows/
        dev.md                        # Branch → PR → merge
        build.md                      # How to build/validate
        release.md                    # How to release/publish
  .github/
    pull_request_template.md          # Status packet + checklist

Each file serves a specific memory function:

File	Memory Function
project.yaml	"What is this project and what matters most?"
permissions.yaml	"What am I allowed to do here?"
onboarding.md	"How do I get started?"
architecture.md	"How is this built and why?"
glossary.md	"What do these domain terms mean?"
decisions/	"What was decided and why?" (accumulates over time)
workflows/	"How do I develop, build, and release?"

Multi-Repo Workspace (Optional)

For teams or individuals with multiple repos, add a registry repo with:

• workspace.yaml — manifest listing all repos (one source of truth)
• bootstrap.sh — clones/pulls all repos on any new machine
• doctor.sh — verifies the toolchain is installed

Any new machine goes from zero to fully synced in 3 commands. See SKILL.md for details.

License

MIT