Continual Learning

Automatically and incrementally keeps AGENTS.md up to date by mining your AI-assisted coding transcripts. Works in both Cursor and Claude Code — teammates using either tool contribute to the same shared memory file.

The plugin watches your conversation turns and periodically extracts recurring corrections and durable workspace facts, then commits them to a dedicated continual-learning branch. Your team curates what lands on main.

Quick install

You're using an AI coding agent — let it handle the setup. Copy the prompt below and paste it into a new conversation.

Claude Code

Install the continual-learning plugin from the GitHub repo samdu/continual-learning-plugin. First ensure bun is installed (bun --version; if missing, install it with curl -fsSL https://bun.sh/install | bash). Then register the repo as a plugin marketplace and install the plugin:

claude plugin marketplace add samdu/continual-learning-plugin
claude plugin install continual-learning

Verify the plugin is installed with claude plugin list.

Cursor

Copy this prompt into a Cursor conversation:

Install the continual-learning plugin from https://github.com/samdu/continual-learning-plugin. First ensure bun is installed (bun --version; if missing, install it with curl -fsSL https://bun.sh/install | bash). Then clone the repo into the Cursor plugin cache so it's available across all projects:

COMMIT=$(git ls-remote https://github.com/samdu/continual-learning-plugin HEAD | cut -f1)
git clone --depth 1 https://github.com/samdu/continual-learning-plugin.git \
  ~/.cursor/plugins/cache/cursor-public/continual-learning/$COMMIT

After cloning, reload the Cursor window (Cmd+Shift+P → "Developer: Reload Window") so Cursor picks up the new plugin.

If your team is on a Cursor Teams or Enterprise plan, an admin can also import this repo as a team marketplace (Dashboard > Settings > Plugins > Import) for one-click installs.

Manual installation steps

Prerequisites

Bun is required — the hook and helper scripts are TypeScript files executed directly via bun run.

curl -fsSL https://bun.sh/install | bash

Note (macOS/zsh): The bun installer adds ~/.bun/bin to your PATH in ~/.zshrc, but hooks run as non-interactive processes that don't source .zshrc. If the hook fails with command not found: bun, add the PATH entry to ~/.zprofile instead:

echo 'export PATH="$HOME/.bun/bin:$PATH"' >> ~/.zprofile

Cursor

Clone the repo into the Cursor plugin cache (user-level, applies to all projects):

COMMIT=$(git ls-remote https://github.com/samdu/continual-learning-plugin HEAD | cut -f1)
git clone --depth 1 https://github.com/samdu/continual-learning-plugin.git \
  ~/.cursor/plugins/cache/cursor-public/continual-learning/$COMMIT

Then reload the Cursor window (Cmd+Shift+P → "Developer: Reload Window").

Team marketplace (Teams/Enterprise plans): An admin can import this GitHub repo via Dashboard > Settings > Plugins > Import instead.

Claude Code

This repo ships a marketplace.json manifest, so it can act as its own single-plugin marketplace. Installation is two steps: register the repo as a marketplace source, then install the plugin from it.

From inside an interactive Claude Code session (slash commands):

/plugin marketplace add samdu/continual-learning-plugin
/plugin install continual-learning

From your terminal (CLI):

claude plugin marketplace add samdu/continual-learning-plugin
claude plugin install continual-learning

Tip: You may see the plugin@marketplace syntax elsewhere (e.g. continual-learning@continual-learning). The @marketplace qualifier is only needed when the same plugin name exists in multiple marketplaces. For most setups you can omit it.

How it works

A stop hook fires after each completed conversation turn. When enough turns have accumulated and enough time has passed, the hook tells the agent to run the continual-learning skill, which:

1. Reads the current AGENTS.md from the continual-learning branch in git (not the local filesystem).
2. Scans transcripts from both Cursor and Claude Code across all worktrees.
3. Processes only new or changed transcripts (incremental index).
4. Extracts high-signal patterns — recurring corrections, durable workspace facts.
5. Commits the updated AGENTS.md to the continual-learning branch and pushes.

Two-tier workflow

Tier 1 (automated): Each skill run commits directly to a long-lived continual-learning branch. No PR, no review friction — learnings accumulate immediately.

Tier 2 (manual curation): The team opens a PR from continual-learning → main whenever they're ready to review and merge. Squash-merge is recommended for clean history.

This keeps the fast path frictionless while giving the team a curation checkpoint before learnings land on main.

Falls back to writing AGENTS.md directly to the workspace root for repos without a remote.

Trigger cadence

Default cadence (after trial window expires):

Parameter	Default
Minimum turns	10
Minimum minutes since last run	120

Trial mode (first 24 hours after install):

Parameter	Default
Minimum turns	3
Minimum minutes since last run	15

Env overrides

All settings can be overridden via environment variables:

• CONTINUAL_LEARNING_MIN_TURNS
• CONTINUAL_LEARNING_MIN_MINUTES
• CONTINUAL_LEARNING_TRIAL_MODE (1/true/yes/on)
• CONTINUAL_LEARNING_TRIAL_MIN_TURNS
• CONTINUAL_LEARNING_TRIAL_MIN_MINUTES
• CONTINUAL_LEARNING_TRIAL_DURATION_MINUTES

Legacy CONTINUOUS_LEARNING_* prefixes also work.

Git worktree support

If you use multiple worktrees of the same repo (common for parallel AI coding sessions), the plugin collapses them to repo level:

• Shared cadence state: The hook resolves the main worktree via git rev-parse --git-common-dir and stores state/index there. Multiple worktrees won't redundantly trigger learning runs.
• Cross-worktree transcript discovery: When the skill runs, it calls git worktree list --porcelain to find all worktree paths, then scans transcript directories for each. Learnings from any worktree are visible everywhere.
• Shared output via git: All worktrees commit to the same continual-learning branch, so learnings from any checkout are visible to every teammate.

Falls back to single-path behavior for non-git workspaces.

State files

Each harness stores its own cadence state and transcript index under the main worktree root (or workspace root if not in a git repo):

Harness	State directory
Cursor	.cursor/hooks/state/
Claude Code	.claude/hooks/state/

AGENTS.md output

The skill maintains two sections:

• ## Learned User Preferences
• ## Learned Workspace Facts

Plain bullet points only. No metadata, no confidence scores.

License

MIT