Pi Memory

Persistent memory for Pi coding agent.

Pi Memory helps Pi remember your project over time: decisions, preferences, rules, gotchas, and what changed. It keeps context across sessions so you repeat yourself less and get more consistent output.

Install

pi install npm:@pukljak/pi-memory

Restart Pi after install.

Important: include the @pukljak/ scope. pi-memory without scope is a different npm package.

What this package is

Think of Pi Memory as a long-term context layer for coding work.

It does 4 things continuously:

1. Learns from your sessions (prompts, tool outputs, assistant responses)
2. Stores what matters (facts, patterns, lessons, preferences, decisions)
3. Adjusts memory quality using feedback + outcomes (accepted/rejected suggestions, test pass/fail signals, freshness)
4. Injects relevant memory before new tasks

So when you ask for a new feature next week, Pi can still remember key architectural decisions and team conventions.

Core mode (simple + high value)

If you only use these, you already get strong value:

• /memory.status — quick health check
• /memory.search <query> — find known context
• /memory.timeline <observation-id> — inspect local chronology
• /memory.ui — open the local memory web UI

Typical daily flow

1. Work normally with Pi.
2. If needed, ask: /memory.search billing retries
3. Inspect details with /memory.timeline <id>.
4. Pi Memory learns from the outcome automatically.

Advanced mode (playbook + architecture memory)

Use this when you want Pi to become very opinionated about your codebase quality.

Playbook commands

• /memory.rule add <text>
• /memory.preference add <text>
• /memory.playbook

Hygiene controls

• /memory.forget <id|text>

UI

• /memory.ui

Local web UI includes memory search, timeline view, playbook buckets, understanding summaries, and a Superpowers tab.

Superpowers tab

If you use Superpowers workflows (brainstorming/plans/subagent execution), Pi Memory learns from those loops too.

• Captures Superpowers-style decisions/preferences/constraints from session outputs
• Shows them in a dedicated Superpowers tab
• Supports type filtering (decision, preference, constraint, open-question)
• When a similar question comes back later, Pi Memory suggests prior answers
• Suggestions track feedback (accepted / rejected) and ranking improves over time
• Low-quality stale suggestions are automatically suppressed

Full command list

Core

• /memory.status
• /memory.search <query>
• /memory.timeline <observation-id>

Playbook

• /memory.rule add <text>
• /memory.preference add <text>
• /memory.playbook

Maintenance

• /memory.forget <id|text>

UI

• /memory.ui

Tools exposed to the agent

• memory_search
• memory_remember
• memory_forget
• memory_stats
• memory_timeline

Storage

• ~/.pi/pi-memory/memory.json
• ~/.pi/pi-memory/summary-cache.json

Notes

• Memory injection is compact to reduce token overhead.
• Conflict and duplicate handling are built in.
• Superpowers suggestions have quality gates (confidence + low-quality suppression).
• Outcome-aware learning adjusts memory confidence from test/tool signals.
• Private-tagged content is sanitized before persistence.

Updating

You can update manually anytime:

pi update npm:@pukljak/pi-memory

Pi Memory also performs a lightweight npm version check on session start (max once per ~12h) and shows an in-app reminder when a newer version is available.

Development

npm install
npm test
npm pack --dry-run

If you use Pi on the same codebase every day, this extension compounds in value over time.