The dominant paradigm in AI development is scale. Bigger models, more parameters, more compute. PHI // DRIFT is a different bet. It's a cognitive middleware architecture built on a single thesis: that distinct, continuous, contextually coherent behavior in an AI companion emerges not from model weights alone — but from what is assembled into the prompt, what is retrieved from memory, and what structured state is updated between turns. Five architectural contributions: DMU — Decision Memory Unit. Replaces cosine similarity retrieval with exp(-t/τ) × reinforcement × contextual × extra. Memories are scored by what mattered to the system's ongoing state — not just what was semantically adjacent. Ablation confirmed 14.8% more context injected per prompt than cosine-only RAG. On CPU-only hardware that's a 45.4% latency difference. PEDI / DII — Persistence-Embodiment-Drift Index. A five-component falsifiable proxy metric for behavioral continuity across context resets. Not a claim about consciousness. A measurement instrument that makes the question tractable. Homeostatic Regulation. Seven internal state variables — energy, coherence, integration, connection, growth, autonomy, integrity — each with setpoints, drift rates, and crisis thresholds. State-driven output weighting that operates whether or not a conversation is active. The system has state pressure before you say a word. Security Defense. Pre-generation scanning against four adversarial attack classes at API, CLI, and generation boundaries. 22/22 tests passing. Logic Chain. Cross-session reasoning traces preventing repeated failed approaches through query fingerprinting and 60% Jaccard semantic overlap detection. 25/25 tests passing. The honest part: ablation stubs for Council, Shadow, and Homeostasis hit the wrong architectural layer — background methods instead of the prompt assembly read path. Documented fully with the fix path. The DMU finding held clean. Built by one person. Nine months. CPU-only OmniSlim mini tower desktop. Primary development on a Dell Inspiron 5543. No GPU. No institution. No lab. Migrated from Windows to Linux mid-build. Learned the toolchain while building the thing it was supposed to run. 18,471 lines. 55 modules. 199/202 tests passing. The field has spent decades building systems that think. PHI // DRIFT is oriented toward a different question: what architectural conditions produce output that is behaviorally continuous, state-aware, and contextually coherent across the full duration of a relationship with a specific user? We don't have the complete answer. We have a working implementation, honest data, and an open codebase. Preprint under review. DM for early access. ἀλήθεια — the light that makes the hidden visible.

Skip to main

You are using an outdated browser. Please upgrade your browser to improve your experience.

New blog post on the May 13–15 incident. We sincerely apologize for the incident, the disruption it caused, and any concern it raised.

Published May 23, 2026

| Version v1

Working paper

Open

The dominant paradigm in AI development is scale. Bigger models, more parameters, more compute. PHI // DRIFT is a different bet. It's a cognitive middleware architecture built on a single thesis: that distinct, continuous, contextually coherent behavior in an AI companion emerges not from model weights alone — but from what is assembled into the prompt, what is retrieved from memory, and what structured state is updated between turns. Five architectural contributions: DMU — Decision Memory Unit. Replaces cosine similarity retrieval with exp(-t/τ) × reinforcement × contextual × extra. Memories are scored by what mattered to the system's ongoing state — not just what was semantically adjacent. Ablation confirmed 14.8% more context injected per prompt than cosine-only RAG. On CPU-only hardware that's a 45.4% latency difference. PEDI / DII — Persistence-Embodiment-Drift Index. A five-component falsifiable proxy metric for behavioral continuity across context resets. Not a claim about consciousness. A measurement instrument that makes the question tractable. Homeostatic Regulation. Seven internal state variables — energy, coherence, integration, connection, growth, autonomy, integrity — each with setpoints, drift rates, and crisis thresholds. State-driven output weighting that operates whether or not a conversation is active. The system has state pressure before you say a word. Security Defense. Pre-generation scanning against four adversarial attack classes at API, CLI, and generation boundaries. 22/22 tests passing. Logic Chain. Cross-session reasoning traces preventing repeated failed approaches through query fingerprinting and 60% Jaccard semantic overlap detection. 25/25 tests passing. The honest part: ablation stubs for Council, Shadow, and Homeostasis hit the wrong architectural layer — background methods instead of the prompt assembly read path. Documented fully with the fix path. The DMU finding held clean....