Sloth Brief is a tool designed to condense accumulated text notes into concise, source-linked action briefs, aiming to reduce rereading effort when resuming tasks. It operates using local Hugging Face Transformers with Qwen2.5-0.5B-Instruct, eliminating the need for paid APIs, cloud uploads, or accounts for inference. The first run downloads model weights, with subsequent inferences using a local cache. A reproducible demo showcases its functionality, processing synthetic notes like reviewing documentation or handling CSV importer failures. The output provides summaries with source attribution, exemplified by the demo's CPU output. Eight automated tests ensure input limits, output formatting, and source references, though they do not guarantee model accuracy. Sloth Brief leverages openly available, Apache-2.0 licensed Qwen2.5-0.5B-Instruct weights, allowing users to inspect the application and replace the model. The tool is deliberately small and CPU-capable, summarizing at most five notes separately per run. The initial version, which requested JSON, was refined to produce one concise summary per note to ensure deterministic source attribution. While source attribution is accurate, the summary's faithfulness to the original note is not guaranteed, requiring user verification. Input files remain unedited, with output generated as a new file, and size limits and source hashes help manage workload. The model lacks execution privileges, preventing malicious notes from running shell commands. The primary benefit is a quick restart checklist, although no specific productivity claims are made. The project was built with AI assistance, and the model demo ran locally on real open weights without using client data.