Skopus

Persistent context for AI coding assistants. Install it, run it, your agent remembers you.

σκοπός (skopos) — Greek for watcher, purpose. A system that gives agents durable scope across sessions.

The problem

Every AI coding assistant forgets you at the end of every session. You teach it your preferences, it forgets. You correct it, it repeats the mistake next week.

The fix

pip install skopus        # or: pipx install skopus
skopus init               # answer 9 questions, everything set up

That's it. Next time you open your AI assistant, it already knows:

• How you work (your non-negotiables, communication style, anti-patterns to avoid)
• What you've corrected (mistakes it won't repeat)
• What you've decided (a knowledge base that grows over time)
• What your code looks like (automatic codebase map via graphify)

Works with Claude Code, Cursor, Codex, Aider, Gemini CLI, and Copilot CLI.

How it works

Skopus writes one file into your project (CLAUDE.md / AGENTS.md / .cursor/rules/) that teaches the AI assistant who you are. The assistant reads it automatically at session start. No re-teaching.

Everything lives in one directory: ~/.skopus/

~/.skopus/
├── charter/     Your rules, non-negotiables, working style
├── memory/      Corrections and wins that compound over time
└── vault/       Your knowledge base (decisions, learnings, sources)

Commands

Setup (run once):

skopus init              # wizard + scaffold + wire current project
skopus link              # wire a different project (if not done during init)

Daily use (inside your AI assistant):

Maintenance:

skopus update            # upgrade + re-install everything
skopus doctor            # health check

Benchmarks

Skopus ships with a benchmark suite that measures whether it actually works. The novel Correction-Persistence benchmark tests: does the agent apply yesterday's corrections to today's tasks?

skopus bench run cp --ablation    # run across 5 lens configurations
skopus bench list                 # see all available benchmarks

Contributing

• New platform adapters: one Python file implementing a 5-method ABC. See skopus/adapters/claude_code.py.
• Benchmark scenarios: real corrections from real sessions. Run /bench-contribute inside Claude Code to generate anonymized scenarios from your feedback.
• Bug reports: github.com/cvalentinhorizontes/skopus/issues

License

MIT — see LICENSE.