MemPalace Autopilot

A Claude Code plugin that gives every session persistent, project-aware memory. At session start it injects the most relevant memories into context; throughout the session it automatically captures conversation deltas into MemPalace on a configurable interval and on every context compaction. Nothing leaks — secrets are scrubbed before anything is stored.

Prerequisites

Requirement	Version
Python	3.10+
Claude Code	latest
MemPalace	`pip install mempalace` (or `pipx install mempalace`)

After installing MemPalace:

mempalace init
mempalace status        # verify it's working
# Optionally write ~/.mempalace/identity.txt with a short personal bio

Install

claude plugin add ./mempalace-autopilot

Migrating from claude-mem

If you have existing memories in claude-mem, bulk-import them before uninstalling:

python3 scripts/migrate_claude_mem.py --dry-run   # preview what will be imported
python3 scripts/migrate_claude_mem.py             # run the actual import
mempalace status && mempalace search "some known topic"
claude plugin remove claude-mem

How It Works

Hook	Trigger	Action
`SessionStart`	Every new chat	Injects ~800 tokens of project-scoped or global memories
`Stop`	After each assistant turn	Every 15 human messages, captures conversation delta
`PreCompact`	Before context compression	Emergency save; blocks until MemPalace finishes
`SessionEnd`	Session close	Captures final delta; cleans up session state

Architecture

mempalace-autopilot/
├── hooks/
│   ├── hook_runner.py          # Single Python entry point for all lifecycle events
│   └── hooks.json              # Claude Code hook configuration
├── scripts/
│   ├── context_generator.py    # Assembles ~800-token memory block at session start
│   └── migrate_claude_mem.py   # One-shot bulk import from claude-mem SQLite DB
├── shared/
│   └── utils.py                # Canonical detect_wing(), scrub_secrets(), SCRUB_PATTERNS
└── tests/                      # 103 unit tests — no network, no MemPalace required

All MemPalace interaction uses subprocess.run(["mempalace", ...]) — the plugin never imports MemPalace directly, so it works regardless of how MemPalace is installed (pip, pipx, venv, system).

Wing Detection

The plugin uses the git repository name as the "wing" (namespace) for stored memories:

Inside a git repo: wing = repo basename (e.g. mempalace-autopilot)
Outside any git repo: wing = general

This means memories from different projects are automatically separated and the most relevant ones are surfaced first at session start.

Security

All content is scrubbed for secrets before being sent to MemPalace. The scrubber targets:

AWS access key IDs (AKIA...)
OpenAI / Stripe secret keys (sk-...)
GitHub tokens (ghp_, gho_, github_pat_)
Bearer tokens
Slack tokens (xoxb-, xoxp-, xoxs-)
Generic api_key=, secret=, token=, password= patterns

Any match is replaced with [REDACTED] before the text leaves the plugin. See SECURITY.md for details on reporting vulnerabilities.

Development

# Run tests (no external services needed)
python3 -m pytest tests/ -v

# Lint
pip install ruff
ruff check .

All tests use unittest.mock — MemPalace and git are fully mocked. The suite runs in under a second.

License

MIT — see LICENSE.