learnlance 🧠🔍

Turn what AI coding agents build into a growing personal knowledge graph.

Demo · Install · Contribute

┌───────────────────────────────────────────────┐
│                                               │
│                YOUR LEARNING GRAPH            │
│                                               │
│    Delta encoding ──── Data synchronization   │
│          │                   │                │
│          └─── Compression ────┘                │
│                  │                            │
│                  └──────── Diffing            │
│                                               │
│    Debouncing ───── Idempotency keys          │
│                                               │
│    Topological sort ─── Graph algorithms      │
│                                               │
└───────────────────────────────────────────────┘

You use AI to write code.

Learnlance watches what gets built and turns that work into concepts you can actually understand and remember.

AI writes the code. Learnlance helps you learn from it.

Demo

learnlance watches your agent — Claude Code, OpenAI Codex, Cursor, GitHub Copilot (CLI / cloud / VS Code), Command Code, Kiro, Gemini CLI, Antigravity, or plain git commit — and after every turn that writes or edits code, it quietly:

1. pulls out the code the agent just wrote,
2. asks an LLM "what transferable concepts could a developer learn from this?" — e.g. "you used a delta function — here's what delta encoding is",
3. merges those concepts into a persistent knowledge graph, and
4. regenerates an interactive HTML graph you can open any time.

Over time you get a browsable map of everything you've picked up while coding — new nodes light up as 🌱 new topics learned.

No API key required. By default it reuses an LLM CLI you're already logged into (claude, gemini, copilot, cursor-agent, or ollama). Or use --in-chat and the agent analyzes its own work — no separate CLI at all.

Why it won't slow you down or break your session

• Analysis runs in a detached background process — your session never waits.
• Every hook is wrapped so a failure is logged and swallowed; it can't interrupt your coding session.
• Turns with no substantive code make no LLM call (no cost, no noise).

Install

pip install learnlance
learnlance setup

setup detects the agents you have and writes their hooks for the current project. By default it analyzes your work with a separate LLM CLI (claude, gemini, copilot, cursor-agent, or ollama).

Using a chat agent (Copilot Chat in VS Code, Cursor, Kiro, Command Code, Antigravity, or Gemini)? Let the agent analyze its own work in the chat instead — no separate CLI to install:

learnlance setup --in-chat

Run setup again in each project you want tracked, then reload your editor and code as usual.

Check what's actually installed and firing:

learnlance doctor

Getting Started

New to LearnLance?

Follow the Getting Started Guide for a step-by-step walkthrough from installation and setup to your first knowledge graph.

Install one agent by hand

learnlance install with no flags is the same as learnlance setup — it detects your agents and configures them all. Pass a flag to configure just one agent, skipping detection:

learnlance install --cursor        # Cursor only
learnlance install --codex         # OpenAI Codex only
learnlance install --copilot       # Copilot CLI / cloud / VS Code Chat
learnlance install --commandcode   # Command Code
learnlance install --kiro          # Kiro
learnlance install --gemini        # Gemini CLI
learnlance install --antigravity   # Antigravity
learnlance install --git           # git post-commit (universal fallback)

Run learnlance install --help for the full list and options.

Supported agents

Each integration is written from the vendor's hook docs. learnlance doctor reports what's configured on disk versus what has actually fired.

Use it

learnlance show      # render + open the knowledge graph in your browser
learnlance list -v   # list learned concepts, with explanations
learnlance stats     # quick counts by category
learnlance help      # every available command

The show view opens with a live loading spinner, then two declutter controls: show related concepts (reveal dimmed umbrella nodes) and a min link strength slider (hide one-off links).

Add a concept the agent missed

learnlance add "debouncing"                 # search the current dir for the topic
learnlance add "topological sort" --path ./src
learnlance add "event sourcing" --force     # add even if it's not in the code

Clear the graph

learnlance clear "delta encoding"   # remove one concept (+ orphaned related nodes)
learnlance clear                    # wipe the graph (asks first; -y to skip)

Per-session recaps are written to ~/.learnlance/insights/<session>.md.

Configuration

learnlance config                            # show current settings
learnlance config --llm-cmd "ollama run llama3"   # use any CLI that reads stdin
learnlance config --cli-model haiku          # model alias for the CLI backend
learnlance config --max-topics 3             # fewer concepts per turn
learnlance config --background off           # run analysis inline (blocks)
learnlance config --disable                  # pause without uninstalling hooks
learnlance config --enable                   # re-enable

Everything lives under ~/.learnlance/. Each project keeps its own graph at projects/<project>-<hash>/graph.json; graph.html is the multi-project dashboard you can open to switch between projects. insights/ holds per-session recaps and learnlance.log holds diagnostics.

Uninstall

learnlance uninstall                 # Claude Code (default)
learnlance uninstall --commandcode   # one specific agent

How it works

learnlance normalizes every agent's hook payload into one CodeEvent, then runs a harness-blind pipeline: insights → knowledge graph → HTML → recap. See ARCHITECTURE.md for the full design.

Zero third-party dependencies by design — hooks must run reliably wherever an agent launches them.

Contributing

LearnLance is actively looking for contributors. You don't need to understand the entire codebase to contribute.

Areas to help

• A coding-agent integration (adapter + installer)
• Knowledge graph algorithms
• Concept extraction
• Graph visualization
• CLI/UX
• Testing
• Documentation
• New learning workflows

Good first contributions

• Add support for another coding agent
• Improve graph visualization
• Add tests for transcript parsing
• Improve Windows compatibility
• Add a CLI command
• Improve concept deduplication
• Improve graph accessibility
• Add documentation or examples

License

MIT — see LICENSE.