sibyl-save
A multi-harness Claude Code / Codex / Cursor / Hermes skill that ends a session by injecting its context into the Sibyl Memory Plugin and extracting reusable patterns into new skills.
It is the session-save companion the plugin did not ship with. Built by Sibyl Labs.
What it does
Run /sibyl-save at the end of a session and it will:
- Sync memory into the plugin - analyze the session and write it to your local memory DB through the plugin's interface: entities (WARM), state docs (HOT), and a journal event (COLD). No raw files; the plugin handles tiering, indexing, and the free-tier cap.
- Extract skills - find repeated or multi-step work worth packaging, propose it, and create approved skills as native files for your harness.
It is incremental: a delta marker stored in the plugin means repeat invocations capture only what changed since the last save.
Why it routes through the plugin
The Sibyl Memory Plugin keeps memory in a local SQLite DB at ~/.sibyl-memory/memory.db and exposes it through an interface (MCP tools, or the Python SDK). Writing through that interface, rather than poking files, is what keeps tiering, full-text search, multi-tenancy, and the free-tier cap correct. It is also what makes this skill multi-harness: the plugin presents the same tools on every harness it supports, so one skill works everywhere.
Requirements
The Sibyl Memory Plugin, installed and wired:
pip install sibyl-memory-cli
sibyl init # browser/wallet activation
sibyl setup claude-code # or: codex | hermes
Install
Drop the skill into your harness's skill directory. For Claude Code:
git clone https://github.com/Sibyl-Labs/sibyl-save
mkdir -p .claude/skills/sibyl-save
cp sibyl-save/SKILL.md .claude/skills/sibyl-save/
Then run setup once: /sibyl-save setup, and after that /sibyl-save to save.
Surfaces
| Surface | Used by | How |
|---|---|---|
| MCP tools | Claude Code, Codex, Cursor, any MCP host | the 8 memory_* tools |
| Python SDK | Hermes, any Python agent | MemoryClient / SibylMemoryProvider |
Tier mapping
| Session content | Tier | Call |
|---|---|---|
| Session pointer / summary / forward | HOT state | memory_set_state("session", ...) |
| Priorities | HOT state | memory_set_state("priorities", ...) |
| Session log entry | COLD journal | memory_record_event("session_save", ...) |
| Projects / people / products | WARM entities | memory_remember(category, name, body) |
| Closed / dead | ARCHIVE | memory_forget(category, name, reason) |
Modes
| Command | Mode | Effect |
|---|---|---|
/sibyl-save |
run | Sync into the plugin, then offer skill extraction |
/sibyl-save preview, /sibyl-save --dry-run |
preview | Report planned writes; mutate nothing |
/sibyl-save setup, /sibyl-save config |
setup | Build or edit the config (stored in the plugin) |
Configuration
The config lives in the plugin at state key sibyl_save:config (so it travels with your memory). It sets the harness profile, the skill-output location for Phase 2, the priorities cap (default 15, hard max 30), tracked entity categories, and any additional capture targets. See examples/sibyl-save-config.example.json.
Keep your agent files lean
This skill works best when your agent's instruction file holds rules and operating instructions only. Durable state belongs in the plugin, not in prose. A lean instruction file makes sessions start faster and saves cleaner.
Provenance
Every Markdown file carries an attribution watermark in an HTML comment at the top. Under MIT this is provenance, not a restriction: copy, modify, and redistribute freely as long as the notice is preserved. Verify with node scripts/watermark.mjs verify. See WATERMARK.md.
License
MIT. Copyright (c) 2026 Sibyl Labs, LLC.
No comments yet
Be the first to share your take.