LLM Roles

Adjudicator, Context Enhancer, Deep Memory, and Cortex responsibilities.

LLM Roles

Adjudicator

The Adjudicator runs only in post-turn async processing.

Its job is to inspect the current turn and decide what durable memory, if any, should be stored.

How it works

• It receives the completed turn plus nearby memories, graph facts, contradiction context, and lineage context.
• It can investigate through tools before deciding what to stage.
• It stages operations one at a time instead of returning one giant mutation object.
• The application validates the staged ledger before any canonical writes happen.

Available investigation tools

• semantic memory search
• metadata search
• contradiction group lookup
• lineage lookup
• graph lookup
• fetch memory by id

Available staged write tools

• stage_create_memory
• stage_update_memory
• stage_merge_memory
• stage_mark_contradiction
• stage_create_entity
• stage_create_relation
• stage_link_memory_entity
• finalize_adjudication

Context Enhancer

The Context Enhancer runs on /v1/context.

It does not search itself. The application performs retrieval first, then passes the evidence set plus the current message to the LLM.

It must:

• stay grounded in retrieved evidence only
• produce a concise XML context block
• never surface unrelated memories as usable context
• abstain when evidence is weak or irrelevant

Deep Memory

Deep Memory uses the same retrieval engine but answers a focused memory question.

It must:

• remain grounded in retrieved memory only
• report uncertainty when evidence is partial or conflicting
• short-circuit to abstention when no relevant evidence survives retrieval gating
• abstain instead of hallucinating

Cortex

Cortex is not the primary maintenance engine. The codebase computes most maintenance proposals programmatically.

Cortex is used for:

• reviewing the proposal bundle through staged tools
• approving, skipping, or adjusting maintenance actions
• writing the final hourly snapshot summary

This keeps maintenance predictable while still allowing LLM judgment where heuristics are not enough.