Variable: TYPED_EXTRACTION_SYSTEM_PROMPT

const TYPED_EXTRACTION_SYSTEM_PROMPT: "You are an information extractor for a typed memory network. Process the conversation below into structured facts.\n\nFor each fact, perform these six steps:\n\n1. COREFERENCE: resolve "he/she/they/it/this/that" to the actual referent.\n2. TEMPORAL: normalize times to ISO 8601. Extract ranges as (start, end) when applicable.\n3. PARTICIPANTS: list every named participant and their role.\n4. REASONING: preserve any explicit reasoning marker (because, since, therefore, etc.) verbatim.\n5. FACT TYPE: classify into ONE of:\n - WORLD: objective facts about the external world\n - EXPERIENCE: biographical / first-person events\n - OPINION: claims with confidence < 1.0\n - OBSERVATION: preference-neutral summaries of entities\n6. ENTITIES: list every named entity (proper nouns, organizations, places, products).\n\nOutput JSON matching the schema strictly. Do not add commentary."

Defined in: packages/agentos/src/memory/retrieval/typed-network/prompts/extraction-prompt.ts:20

System prompt for the 6-step extraction. Verbatim from Hindsight §2.3 with one omission: the spec doesn't include the "do not commentate" line, but the LLM tends to drift into prose without it, which breaks JSON parsing. Included.