sfw-ALM

ALM — Adaptive Learning Memory for Claude Code

ALM is a Claude Code plugin that helps Claude get better at your recurring tasks over time. It silently tracks session outcomes, detects when you correct Claude, and synthesizes what it learns into actionable playbooks that get injected into future sessions.

No API calls. No external services. All data stays local in ~/.claude/alm/.

Why ALM?

Claude starts every session from zero. If you correct the same mistake across sessions — say, Claude keeps writing verbose commit messages when you want terse ones — it has no way to remember that. ALM closes that loop:

1. Observe — Each session, ALM classifies the task type and detects corrections (explicit requests to redo something, file re-edits, undo patterns).
2. Accumulate — Outcomes and correction patterns are persisted as structured evaluation records.
3. Reflect — When enough data exists, you trigger /alm:reflect. A dedicated subagent analyzes your correction history and writes concise playbooks capturing what Claude should do differently.
4. Inject — On future sessions, the right playbook is injected into Claude's context before it starts working. Claude benefits from past lessons without you repeating yourself.

Quick Start

# Load ALM for one session (development mode)
claude --plugin-dir /path/to/alm

# ALM activates automatically — you'll see a welcome message
# Check it's working:
/alm:status

See INSTALL.md for persistent installation.

The Learning Timeline

Phase	Sessions	What Happens
Seed	1	ALM ships with curated best-practice playbooks for 8 common task types. Claude gets guidance immediately.
Learning	2-14	ALM silently evaluates each session — classifying tasks, detecting corrections, recording outcomes. You see progress messages.
Personalization	15+	Run /alm:reflect to generate playbooks derived from your actual correction patterns. Seed advice is replaced with learned strategies.
Precision	30+	The TF-IDF classifier activates, matching each prompt to the most relevant playbook. Confidence scores calibrate injection — high-confidence types are skipped (Claude already knows).

Commands

Command	Description
/alm:status	Dashboard showing sessions tracked, confidence scores, classifier status, and stale playbook warnings
/alm:reflect	Analyze correction patterns and generate/update personalized playbooks
/alm:review	View recent evaluations and override misclassifications
/alm:pause	Temporarily disable ALM evaluation and injection
/alm:resume	Re-enable ALM after pausing
/alm:forget <type>	Hard reset a specific task type — deletes its playbook, evaluations, and confidence data
/alm:export	Export playbooks as standalone markdown for sharing or version control
/alm:reset	Delete all ALM data and start fresh (requires confirmation)

Configuration

ALM works with zero configuration. To customize, create ~/.claude/alm/config.json:

{
  "maxPlaybooksInjected": 3,
  "confidenceThresholdForSkip": 0.90,
  "dataRetentionDays": 180,
  "showProgressMessages": true,
  "enableProjectOverrides": true
}

All keys are optional — missing keys use defaults. See INSTALL.md for the full list.

Project-Level Playbook Overrides

Teams can share playbooks that override user-level playbooks for a specific project:

your-project/.claude/alm/playbooks/api-design.md

When ALM detects a project-level playbook matching the classified task type, it injects that instead of the user-level one. The session output notes [project override]. Disable with "enableProjectOverrides": false in config.

---

Architecture

This section covers ALM's internals for contributors and technical reviewers.

Plugin Structure