Building Mneme: An Autonomous AI Content Factory

Part 1

The Spark and the Spine: Architecture and Why Local

Core stack, multi-LLM router, persona system with LoRA lineage, and how MCP turned reasoning into real work.

~8 min read · with overview diagram + 90s demo

Part 2

From Lessons to E‑books: Persona Tuning and LoRA‑Driven Quality

Why Lesson Creator fell short and how E‑book Creator improved quality with persona iteration and contracts.

~7 min read

Part 3

Weller Davis: Voice Cloning, Audiobooks, and a Real Publishing Loop

Chatterbox voice cloning, “Virtual Me,” and the pragmatic realities of validation vs. automation.

~7 min read

Part 4

Images at Scale: Picaso, Fooocus → ComfyUI, and Surviving the Chaos

Network/firewall issues, model management, quality gates with Vision LLM, and dynamic LoRA loading.

~8 min read

Part 5

The Music Creator Odyssey: YuE on Consumer Hardware

FlashAttention2 to SDPA, PyTorch/CUDA pinning, xcodec checkpoints, and the 13-second breakthrough.

~9 min read

Part 6

Agentic Coding: Mneme’s Plan → Implement → Validate → Fix Loop

Incremental dev plans, runtime tests, and the roadmap to SOTA with reflection and test‑first strategies.

~9 min read

Part 7

The System Pattern Library: Scheduler, Filesystem, Recovery

Reusable patterns for long-running workflows, artifacts, retries, and sleep cycles.

~7 min read

Part 8

The Bigger Vision: Local, Specialized Models and a Fairer AI Future

Personal AI at the edge, specialized personas, and the case for democratized AI capabilities.

~6 min read

About this series

Mneme is a production AI system built by Dave Wheeler (aka Weller Davis). It features a multi‑persona architecture with LoRA fine‑tuning, an MCP tool layer for real‑world actions, ComfyUI pipelines for image/music, and a family of autonomous creator modules.