Skip to the content.

LLM Context Layer Model

LLMs are stateless. Every turn, the agent runtime assembles a complete context payload and sends it as one request. The payload has a consistent layered structure.

Layer Stack

Layer Content Position Injection
L0 System prompt (identity, tool schemas, extension instructions, skills list) Front Every turn (runtime)
L1 Global hints (global .goosehints, MOIM) After L0 Every turn (runtime)
L2 Project hints (AGENTS.md, SOUL.md) After L1 Every turn (runtime)
L3 Conversation history Middle Every turn (grows, compacts)
L4 Turn context (timestamp, working directory) Before user message Every turn (runtime)
L5 Current user message Last Every turn (user)

All layers are sent every turn. The difference between layers is position, not presence.

Attention Zones

The primacy and recency effects are well-replicated across models. The degree of middle decay is model-specific and has been reduced in modern long-context models.

Key Properties

Harvested From

Project Source Entry
context-eng Session: Context Landscape Analysis Full analysis of Goose context assembly, empirical observation of layer positions

Implications