Dual-harness continual learning plugin — keeps AGENTS.md up to date from transcripts
npx claudepluginhub samdu/continual-learning-pluginIncrementally learns durable user preferences and workspace facts from transcript changes and keeps AGENTS.md up to date. Works in both Cursor and Claude Code.
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.
You're using an AI coding agent — let it handle the setup. Copy the prompt below and paste it into a new conversation.
Install the continual-learning plugin from the GitHub repo
samdu/continual-learning-plugin. First ensure bun is installed (bun --version; if missing, install it withcurl -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-learningVerify the plugin is installed with
claude plugin list.
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 withcurl -fsSL https://bun.sh/install | bash). Then clone the repo into the Cursor plugin cache so it's available across all projects:COMMIT=$(git ls-remote https://github.com/samdu/continual-learning-plugin HEAD | cut -f1) git clone --depth 1 https://github.com/samdu/continual-learning-plugin.git \ ~/.cursor/plugins/cache/cursor-public/continual-learning/$COMMITAfter 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.
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/binto yourPATHin~/.zshrc, but hooks run as non-interactive processes that don't source.zshrc. If the hook fails withcommand not found: bun, add the PATH entry to~/.zprofileinstead:echo 'export PATH="$HOME/.bun/bin:$PATH"' >> ~/.zprofile
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.
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@marketplacesyntax elsewhere (e.g.continual-learning@continual-learning). The@marketplacequalifier is only needed when the same plugin name exists in multiple marketplaces. For most setups you can omit it.
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:
AGENTS.md from the continual-learning branch in git (not the local filesystem).AGENTS.md to the continual-learning branch and pushes.Tier 1 (automated): Each skill run commits directly to a long-lived
continual-learning branch. No PR, no review friction — learnings accumulate immediately.