Project Thalamus Technical Whitepaper: An Authorship Gate Between Clients and Language Models

Abstract

Language models recompute answers from scratch on every request. Asking a question whose answer was already reviewed and approved by a human still burns full compute time and introduces subtle phrasing drift. Standard response caches attempt to fix this, but because they cache automatically, they regularly end up storing hallucinations and serving them indefinitely. Project Thalamus solves this by placing an explicit authorship gate between client applications and the underlying language model. Rather than treating memory as a probabilistic guess, Thalamus enforces strict operational rules. If a question matches a human-verified entry in a local JSON brain through an exact string match or a MiniLM cosine similarity of 0.72 or higher, Thalamus returns the exact stored text in milliseconds without calling the model. If no approved answer exists, the request routes directly to the model every single time. Any request carrying function definitions or a tools payload bypasses memory completely to keep active agent loops working smoothly. Models cannot write to memory on their own, as learning is strictly opt-in through Learner Mode to ensure unreviewed model outputs never leak into stored facts. Version 0.8 details the verified system architecture, which relies on lightweight local JSON files rather than complex database services. It covers live-tested endpoints, including OpenAI-compatible completions and Model Context Protocol (MCP) transports over Streamable HTTP and stdio, alongside latency benchmarks against standard RAG pipelines. Every claim in the whitepaper is backed by live trace audits and a green 127-test suite.