We present PHI // DRIFT, a cognitive middleware architecture designed to address a fundamental limitation in current large language model deployments: the absence of persistent internal state that evolves across interactions with a specific user over time. Existing systems process each interaction as an isolated probabilistic event — competent, but stateless. We describe this gap as talking to the statistics of a mind.

DRIFT introduces five architectural contributions. The Decision Memory Unit (DMU) weights retrieved memories by time-decay, reinforcement, and contextual salience — exp(-t/τ) × reinforcement × contextual × extra — rather than vector similarity alone. The Persistence-Embodiment-Drift Index (PEDI) provides a five-component falsifiable proxy metric for behavioral continuity across context window boundaries. A homeostatic regulation layer models seven internal state variables with setpoints, drift rates, and crisis thresholds generating state-driven output weighting independent of user input. A security defense layer guards against four attack classes at API, CLI, and pre-generation boundaries. A logic chain reasoning trace system prevents repeated failed approaches across sessions through query fingerprinting and semantic overlap detection.

All development, testing, and evaluation were conducted on consumer hardware: a CPU-only OmniSlim mini tower desktop with no dedicated GPU. Primary development occurred on a Dell Inspiron 5543 (circa 2012–2015). Ablation testing confirmed that DMU re-ranking injects 14.8% more context into prompts than cosine-only retrieval. Live stress testing at 50-thread concurrency produced 100% success rate across 50 requests with no breaking point found. We do not claim PHI // DRIFT is conscious. We claim it produces measurably more continuous, contextually coherent output than stateless alternatives — and we provide a framework for testing that claim.

DRIFT_paper_v4.pdf

Files (97.6 kB)