persistent-memory
Scannednpx machina-cli add skill a5c-ai/babysitter/persistent-memory --openclawpersistent-memory
You are persistent-memory -- the cross-session knowledge persistence skill for Pilot Shell.
Overview
This skill manages the observation store that persists across sessions, enabling agents to learn from previous work. Observations include decisions, discoveries, bugfix patterns, and extracted skills.
Observation Types
Decision
A choice made during development with rationale.
{ "type": "decision", "content": "Use Redux Toolkit over raw Redux", "rationale": "RTK reduces boilerplate by 60%", "tags": ["architecture", "state-management"] }
Discovery
A codebase insight or pattern found during work.
{ "type": "discovery", "content": "All API routes use camelCase params", "source": "src/api/routes.ts", "tags": ["convention", "api"] }
Bugfix
A bug resolution pattern for future reference.
{ "type": "bugfix", "content": "Race condition in WebSocket reconnect", "rootCause": "Missing mutex on connection state", "fix": "Added connection state lock", "tags": ["concurrency", "websocket"] }
Storage
Observations are stored in .pilot-shell/memory/ as JSON files, one per session:
.pilot-shell/memory/
2026-03-02-session-a1b2c3.json
2026-03-01-session-d4e5f6.json
Retrieval
Search by:
- Tags: Filter by tag match
- Full text: Search observation content
- Type: Filter by decision/discovery/bugfix
- Recency: Prefer recent observations
- Relevance: Score against current task description
Auto-Triggers
/learntriggered at context thresholds- Session end captures final observations
- Breakpoints can trigger observation capture
Source
git clone https://github.com/a5c-ai/babysitter/blob/main/plugins/babysitter/skills/babysit/process/methodologies/pilot-shell/skills/persistent-memory/SKILL.mdView on GitHub Overview
Persistent-memory stores observations across sessions, enabling agents to learn from decisions, discoveries, and bugfix patterns. It makes these insights searchable via tags and relevance scoring so past work informs current tasks.
How This Skill Works
Observations are written to .pilot-shell/memory/ as JSON files (one per session). Retrieval supports filtering by tags, full-text search, type (decision/discovery/bugfix), recency, and relevance scoring against the current task. Auto-triggers include /learn at context thresholds, end-of-session captures, and breakpoint-driven observation logging.
When to Use It
- You need to reuse a past debugging pattern from a prior session.
- You want to trace architectural decisions and the rationale behind them.
- You’re recalling code conventions or discovery patterns across projects.
- You’re starting a new task and want to surface relevant prior observations.
- You need to review recent work to avoid duplicating effort.
Quick Start
- Step 1: Record observations as JSON objects with type (decision/discovery/bugfix), content, and tags.
- Step 2: Ensure each session writes to .pilot-shell/memory/ (one file per session).
- Step 3: Retrieve using tags, full-text search, type, recency, or relevance to surface relevant history.
Best Practices
- Tag observations consistently using standardized tags (e.g., architecture, api, concurrency).
- Record the type (decision/discovery/bugfix) plus any rationale, rootCause, and fix when applicable.
- Include source references (file paths or commits) to provide context for discoveries and bugfixes.
- Review memory before major tasks to surface relevant past patterns and decisions.
- Periodically prune or archive outdated observations to keep memory focused and performant.
Example Use Cases
- Decision: Use Redux Toolkit over raw Redux; rationale: reduces boilerplate by 60%; tags: architecture, state-management.
- Discovery: All API routes use camelCase params; source: src/api/routes.ts; tags: convention, api.
- Bugfix: Race condition in WebSocket reconnect; rootCause: missing mutex on connection state; fix: added connection state lock; tags: concurrency, websocket.
- Storage: Observations saved under .pilot-shell/memory/ with per-session JSON files like 2026-03-02-session-a1b2c3.json.
- Auto-trigger: /learn fires when context thresholds are met, capturing final observations at session end.
