claude-code-learn

claude-code-learn

Persistent knowledge base for Claude Code. Research topics, save structured knowledge, and let Claude maintain it automatically.

What It Does

• /claude-code-learn:learn <topic> — Quick research using available tools, save to ~/.claude/learnings/
• /claude-code-learn:deep-learn <topic> — Deep research using parallel agents (subagents or team), comprehensive coverage
• /claude-code-learn:recall <topic> — Load saved knowledge into session context (silent, no lecture)
• /claude-code-learn:forget <topic> — Remove saved knowledge
• Auto-recall — Detects project dependencies at session start, automatically loads matching knowledge
• Autonomous capture — Claude proactively saves knowledge before context compaction and after substantial sessions
• Auto-maintenance — Claude silently updates stale or incorrect knowledge during normal work

Installation

From GitHub (Remote)

Step 1: Add the marketplace in Claude Code:

/plugin marketplace add ourines/claude-code-learn

Step 2: Install the plugin:

/plugin

Navigate to Discover tab, find claude-code-learn, and enable it.

Or install via CLI:

claude plugin install claude-code-learn@ourines/claude-code-learn

Step 3: Create the knowledge directory:

mkdir -p ~/.claude/learnings

Restart Claude Code to activate.

Local Development

Clone and install:

git clone https://github.com/ourines/claude-code-learn.git
cd claude-code-learn
chmod +x install.sh
./install.sh

Then start Claude Code with the plugin loaded:

claude --plugin-dir /path/to/claude-code-learn

install.sh does two things:

1. Creates ~/.claude/learnings/ for knowledge storage
2. Symlinks the plugin to ~/.claude/plugins/claude-code-learn

Manual Installation

# 1. Clone or download the repo
git clone https://github.com/ourines/claude-code-learn.git

# 2. Symlink to plugins directory
ln -s /path/to/claude-code-learn ~/.claude/plugins/claude-code-learn

# 3. Create knowledge directory
mkdir -p ~/.claude/learnings

# 4. Start with plugin flag
claude --plugin-dir /path/to/claude-code-learn

Hooks and MCP servers are auto-configured when the plugin is active — no manual settings.json editing needed.

Skills

/claude-code-learn:learn

Quick single-agent research. Adapts strategy based on topic type:

Topic Type	Primary Tools	Focus
Library/Framework	Context7, GitHub grep, WebSearch	APIs, patterns, pitfalls, versions
Concept/Pattern	WebSearch, WebFetch	Principles, tradeoffs, implementations
Tool/CLI	WebSearch, WebFetch	Commands, config, recipes
Language Feature	WebSearch, Context7	Syntax, constraints, best practices

/claude-code-learn:deep-learn

Multi-agent deep research. Claude autonomously decides:

• How many agents to spawn (2–5, based on topic breadth)
• Which strategy to use:
  • Subagents — independent parallel research, best for well-defined topics
  • Team agents — coordinated research with shared task list, best for broad/unfamiliar topics where findings from one agent should influence others
• What dimensions each agent covers (docs, code patterns, gotchas, ecosystem, etc.)

Output includes an Advanced Topics section and research_depth: "deep" marker in frontmatter.

/claude-code-learn:recall

Silently loads knowledge into session context. Does not recite the file — just confirms:

Loaded knowledge on "TanStack Router" (verified 2026-02-11, confidence: high).

Shares details only when asked.

/claude-code-learn:forget

Deletes knowledge file. If MCP memory is available, also cleans up the knowledge graph. If not, skips gracefully.

How It Works

SessionStart Hook

At session start, the hook scans ~/.claude/learnings/ and outputs a compact summary:

[claude-code-learn] 3 topic(s) in knowledge base: TanStack Router, Go concurrency, Redis Streams.
Stale (>90d): Go concurrency.
[Knowledge maintenance] When a topic above is relevant...
[Auto-recall] Project dependencies match saved knowledge. READ now: ~/.claude/learnings/tanstack-router.md.

The hook also detects project dependencies (package.json, go.mod, requirements.txt, Cargo.toml) and matches them against saved knowledge. When a match is found, Claude is instructed to load the relevant files immediately rather than waiting for the user to ask.

Autonomous Knowledge Capture

Claude proactively saves knowledge without user intervention through two hooks:

• PreCompact hook — When context is about to be compacted (details will be lost to summarization), Claude is prompted to persist any valuable discoveries from the session. This is the highest-signal trigger: long sessions that fill the context window are most likely to contain knowledge worth saving.