Research

I investigate model authority, memory, and constrained inference through code and reproducible experiments, with evidence and limits for each artifact.

Published software and evidence

active

emotional-memory

Memory layer for LLM systems with affective state encoding, a PyPI package, Zenodo DOI, and reproducible benchmarks against Mem0, LangMem, and Letta.

Research question: When does affect-conditioned retrieval help an LLM memory layer?
Limits: The measured advantage is limited to mood-congruent, affect-discriminative recall with affect supplied at encoding. General QA, automatic appraisal, and several external benchmarks do not show the same advantage. The DOI archives software; it is not a peer-reviewed paper.

Project details Inspect the evidence Archived software (DOI)

Reference implementations

active

reasoning-kernel

A Python reference implementation of a reasoning kernel that separates untrusted model text from authorized effects using capability-based control and taint tracking.

Research question: How can untrusted model output be separated from authorized effects?
Limits: The implementation enforces a mediation and authorization topology. Conformance does not establish that a host policy is safe. It is a reference implementation, not a security certification or an independent audit.

Project details Inspect the evidence

Experiments

active

affective-fly

Interpretable affect source for emotional-memory, built from a reduced Drosophila mushroom-body circuit with persistent mood, approach/avoid decisions, journal replay, and committed benchmarks.

Research question: Can a reduced circuit supply an inspectable affect signal and persistent mood to memory?
Limits: The circuit and emotion labels are computational models, not claims about fly experience. Without a real connectome, the circuit weights are random; mood time constants remain hypotheses pending host-session validation.

Project details Inspect the evidence

active

xllama

Local LLM chat and Stable-Diffusion image generation on Xbox Series S|X in UWP development mode, with ONNX Runtime GenAI and DirectML routed per workload.

Research question: How do local inference workloads fit the memory and platform constraints of Xbox developer mode?
Limits: This is developer-mode and LAN research, not a retail Xbox application or a public inference service. A recorded demo does not establish throughput across models or hardware.

Project details Inspect the evidence

Architecture topics and notes

research note

Agent runtime concepts

Execution loops, role boundaries, termination rules, persistent state, and runtime-level failure handling.

#runtime#agents#orchestration

active experiment

Memory models for LLM systems

Comparing episodic, semantic, compressed, and affective memory for agents that need continuity.

#memory#retrieval#state

architecture note

MCP routing and tool boundaries

How tool catalogs, permissions, connector identity, and protocol adapters shape reliable agent execution.

#MCP#tools#protocols

experiment

LLM inference on constrained hardware

Running local LLM chat and image generation on the Xbox Series S|X via UWP dev mode — ONNX Runtime GenAI plus DirectML, routed per workload under tight platform limits.

#ONNX Runtime#DirectML#on-device

xllama

Read the engineering articles