A logic engine that sends only changed facts between conversation turns instead of replaying full history. Extracts structured predicates from natural language, runs backward-chaining inference to find goal-reachable facts, and returns token-optimized deltas. Exposes `/context` for initial reduction, `/memory/context` for goal-driven queries, and `/context/diff` for incremental updates. Built on Hexastore indexing with truth maintenance, so retracting a fact auto-retracts downstream inferences. Works as an MCP server, REST API, or SDK. Useful when you're burning tokens on repeated tool outputs or state replays and need deterministic, reproducible compression without vector search or summarization drift.
claude mcp add --transport stdio ai.nocturnus-logic-server uvx logic-server