skopus 0.1.4

Persistent four-lens context for AI coding assistants — charter, memory, vault, graph. One install, multi-agent, benchmark-driven.

Skopus

Persistent four-lens context for AI coding assistants. Charter, memory, vault, graph — one install, multi-agent, benchmark-driven.

σκοπός (skopos) — Greek for watcher, lookout, target, purpose. The root of scope, telescope, episcopal. A system that gives agents durable scope across sessions.

---

The problem

Every AI coding assistant — Claude Code, Cursor, Codex, Aider, Gemini CLI, Copilot CLI — loses context at the end of every session. You teach them your preferences, they forget. You correct them, they repeat the mistake next week. Your hard-won lessons evaporate into chat history.

The few persistent-memory systems that exist (claude-mem, Mem0, MemPalace, OpenAI memory) record conversations but don't encode how you work — the non-negotiables, the drift log, the anti-rationalization rules, the "do this, not that" patterns that actually make a collaboration compound. Meanwhile, structural knowledge about a codebase gets rediscovered every session via grep because nothing persists the map.

The result: agents that are smart per-message but stupid across sessions, and humans who spend half their time re-teaching.

The promise

A unified four-lens context system any AI coding assistant can load at session start:

1. Charter — how you work together (non-negotiables, anti-rationalization table, drift log)
2. Memory — what happened before (feedback, corrections, project state)
3. Vault — what you decided and learned (narrative wiki, Karpathy /raw pattern)
4. Graph — what the code looks like (via graphify)

One install. Works with 6+ agents. Ships with a benchmark suite that proves it works.

Quickstart

# Install (pick one)
pip install skopus             # if pip works on your system
pipx install skopus            # on Ubuntu/Debian (recommended — handles venv for you)

# Set up
skopus init                    # interactive wizard (10 questions, ~5 min)
cd my-project && skopus link   # wire the current project to your charter + vault
skopus doctor                  # health check all four lenses

# Later: upgrade to latest
skopus update                  # upgrades skopus + re-installs graphify + vault commands

Note: init, link, update, doctor are subcommands of skopus — not separate packages. Don't try to pip install update or pipx install init. Just type skopus <command>.

What ships at v0.0.3 (alpha)

• ✅ Charter templates — high-level CLAUDE.md, full workflow_partnership.md, user_profile.md
• ✅ Memory scaffold — MEMORY.md index, feedback/project templates, 6 seed profiles
• ✅ Vault scaffold — Karpathy raw/wiki/output layout with /ingest, /compile, /query, /lint, /wiki slash commands
• ✅ Interactive wizard — 10-question personalization flow (+ --non-interactive for CI)
• ✅ Non-destructive init — re-running skopus init preserves user edits by default; --force to overwrite
• ✅ Six platform adapters — Claude Code, Cursor, Codex, Aider, Gemini CLI, Copilot CLI. All idempotent with automatic backup.
• ✅ Graphify integration — hard dependency, automatic wiring of graphify's PreToolUse hook + git post-commit hook, consolidation of graphify's block into .claude/CLAUDE.md
• ✅ skopus charter evolve — session-end reflection loop. Three-question interactive prompt captures validated calls, drifts, and new rules into feedback memory and the charter's drift log. The mechanism that makes the charter compound.
• ✅ skopus doctor — health check across all four lenses plus linked projects
• 🚧 Benchmark harness — LongMemEval, LoCoMo, MSC, RULER, Correction-Persistence — planned for v0.1.0

Supported platforms (v0.0.3)

Platform	Context file	Detection
Claude Code	.claude/CLAUDE.md (preferred) or root CLAUDE.md	~/.claude/
Cursor	.cursor/rules/skopus.mdc (alwaysApply: true)	cursor binary or ~/.cursor/
Codex (OpenAI)	AGENTS.md	codex binary or ~/.codex/
Aider	AGENTS.md	aider binary or ~/.aider.conf.yml
Gemini CLI	GEMINI.md	gemini binary or ~/.gemini/
Copilot CLI	AGENTS.md	gh / copilot binary or ~/.copilot/

See docs/DESIGN.md for the full spec and roadmap.

The benchmark pillar

Skopus is designed to be measurable. The charter's core non-negotiable — evidence over assumption — is applied reflexively to the project itself. Every PR that touches the charter templates, adapter wiring, or wizard flow must move a benchmark number or explain why it's orthogonal.

At v0.1.0, the skopus bench run harness will run:

Benchmark	What it tests
LongMemEval (Wu et al. 2024)	6 memory abilities: single-session, multi-session, knowledge update, temporal reasoning, explicit/implicit refs
LoCoMo (Google 2024)	Long multi-session conversations
MSC (Facebook 2021)	Persona consistency across sessions
RULER (NVIDIA 2024)	Long-context retrieval, up to 128K ctx
Skopus Correction-Persistence (novel)	Does the agent apply yesterday's corrections to today's tasks?

The ablation mode measures the additive contribution of each lens:

skopus bench run all --ablation --agent claude-code
# Runs vanilla / +charter / +memory / +vault / +graph — shows delta per lens

Philosophy

Skopus combines three existing patterns into one coherent system:

• Karpathy's LLM Knowledge Base (tweet, gist) — the raw/wiki/output three-folder split with Ingest/Query/Lint operations, framed as a modern Vannevar Bush Memex.
• The Partnership Charter — evidence over assumption, premium quality, anti-rationalization tables, drift logs. A meta-workflow layer most agent setups don't have.
• Graphify (safishamsi/graphify) — automatic structural knowledge graph extraction from any codebase with honest audit trails (EXTRACTED / INFERRED / AMBIGUOUS).

Nothing here is new on its own. The contribution is the coherent integration plus the benchmark commitment to prove it works.

Contributing

New platform adapters welcome. Each adapter is a single Python file implementing a 5-method ABC (detect, install, uninstall, status, session_end_hook). See skopus/adapters/base.py and skopus/adapters/claude_code.py for the reference implementation.

Benchmark dataset contributions welcome — especially for the Correction-Persistence benchmark, which ships with 100+ scenarios in bench/correction_persistence/dataset.json.

License

MIT — see LICENSE.