Tiered Context-Loading System is the three-level architecture (L0, L1, L2) that the 📝MythOS MCP uses to load augmentation context progressively — small always-on core, deeper context on demand.
The tiered system separates what an agent must know to act from what it can fetch when needed. L0 and L1 load automatically at session start via get_context, delivering identity memos (Soul, Style) and personal context memos (Human, Memory). L2 is opt-in, loaded through load_context_tier, and carries extended context memos tagged #mythos-mcp-context plus the full template catalog. The pattern keeps the default payload lean while making richer context available the moment an agent needs it.
Key Facts
- Category: MCP context architecture
- Origin: 2026, MythOS
- Home:
src/lib/augmentation/constants.ts - Tool surface:
get_context(auto-loads L0+L1),load_context_tier(fetches L2) - Tiers: L0 Identity, L1 Personal, L2 Extended
- Notable for: progressive loading that keeps always-on payloads small while preserving discovery of deeper context
How It Works
- L0 Identity — Soul and Style memos. The immutable core: who the user is, how they write, the values guiding collaboration. Always loaded.
- L1 Personal — Human and Memory memos. Situational context: current projects, confirmed approaches, corrections, preferences. Always loaded alongside L0.
- L2 Extended — memos tagged
#mythos-mcp-contextplus the template catalog. Domain-specific context and scaffolds loaded on demand viaload_context_tier(tier: "L2"). - Budget accounting —
get_contextreports token counts per tier and advertises the cost of loading L2, letting the agent decide whether deeper context is worth fetching.
Why It Matters
The tiered split resolves a tension every MCP faces: agents need enough context to act coherently, but loading everything at session start burns tokens on material most conversations never touch. Tiering promotes the smallest sufficient context to always-on status and defers the rest behind an explicit fetch. The agent still knows L2 exists — it just pays the token cost only when the work calls for it.
The design also makes the context surface legible to the agent itself. L0 is identity, L1 is personal, L2 is extended. When something feels missing, the agent knows exactly where to look and which tool to call.
FAQ
What loads automatically when a session starts?
L0 (Soul + Style) and L1 (Human + Memory) load automatically through get_context. The agent receives identity, formatting rules, personal context, and confirmed collaboration patterns before acting.
When should an agent call load_context_tier?
When a task references extended context — project-specific memos tagged #mythos-mcp-context, or the full template catalog. get_context reports what's available in L2 and its token cost; the agent decides whether to fetch.
Why aren't full templates always loaded?
Template bodies are expensive. Keeping bodies behind L2 preserves a lean always-on context. A lightweight template index — names plus one-line descriptions — can live in L1 to preserve discovery without paying the full content cost every session.
Can new tiers be added?
Yes. The tier definitions in constants.ts are additive. New tiers extend the system as long as they respect the pattern: L0 is identity, L1 is personal, higher tiers are opt-in extended context.
Related
- 📝MythOS — the platform this architecture runs inside