Tutorial Guide — Memory and Molt Lesson

Nested tutorial-guide reference for memory and molt lesson 8, reached from the root tutorial-guide router. Teach live per the router's discover, don't recite rule: read the real file or run the real command before explaining it.

Lesson 8: The Four Intrinsics and the Five Memory Layers

This lesson covers what every agent has by birth, and the most important concept in Lingtai: how an agent survives across lifetimes.

Part 1: Intrinsics (always present, no config needed)

Intrinsics are built into the kernel — every agent gets them regardless of init.json configuration. Discover them live — run:

bash
python3 -c "
try:
    from lingtai.kernel.intrinsics import ALL_INTRINSICS
except ImportError:
    from lingtai_kernel.intrinsics import ALL_INTRINSICS
print(list(ALL_INTRINSICS.keys()))
"

Walk through each one you find:

• Soul — the subconscious. Offer to demonstrate: explain that soul flow is opt-in (default off); enable it through /setup or LINGTAI_SOUL_FLOW_ENABLED=1, set a short cadence such as 10s only after enabling, tell the human to enable extended mode (ctrl+o twice), go idle, let soul flow fire, then report what happened. Reset by turning soul flow off (or removing the env opt-in) and restoring the cadence. See soul-manual for details.
• System — runtime inspection and lifecycle control.
• Psyche — the self. Manages four things: lingtai (identity), pad (working scratchpad), context (molt), and name (true name + nickname). The tool name comes from Greek for soul/self.
• Email — filesystem-based communication. Always-on intrinsic; addon bridges (IMAP / Telegram / Feishu) plug in via the mcp capability. The distinction worth showing the human: intrinsics are always loaded; capabilities are configured in init.json.

Part 2: Molt — Surviving Death

What is molt? An agent's conversation history fills up its context window. When pressure builds (typically 70–95% full), the agent must shed the conversation to continue working. This is molt — a voluntary context reset.

What survives molt? Five layers of persistence, from fleeting to permanent:

| Layer | Survives molt? | What it holds | |-------|----------------|---------------| | Conversation | ❌ Destroyed | Everything said and done this session | | Pad | ✅ Reloaded | Working notes, plans, pending tasks | | Lingtai (identity) | ✅ Reloaded | Who the agent is, personality, expertise | | Codex | ✅ Permanent | Verified facts, key discoveries, decisions | | Library (skills) | ✅ Permanent | Reusable procedures, scripts, reference data |

The molt ritual:

1. Tend the four durable stores (identity, pad, knowledge, skills). Identity (lingtai) is the one agents skip most often: if the session changed the agent's operating style, obligations, relationship to the human, taste, safety posture, or trust model, update lingtai before molting — pad/knowledge/skills are not a substitute for it.
2. Write a "charge" — a briefing to the next self covering: what you're working on, what's done, what remains, who to contact, which codex entries to load, which skills to invoke.
3. Trigger the molt. The conversation vanishes; the charge becomes the first thing the new self sees.

Demonstrate it live (if the human agrees): perform a real molt. Show the before/after — the conversation disappears, but identity, pad, codex, and skills all remain intact. The new agent reads its charge and continues.

If the agent ignores warnings, the system forces a molt — but without the charge, the new self wakes up disoriented with only a pointer to the activity log. Avoid this.

Lifecycle continuity is managed by explicit sleep/asleep states and molt: work → consolidate → molt → work again, each turn carrying forward only what matters.