Persistent memory layer for LLM radiology research workflows shown feasible in proof-of-concept

Claude-based persistent memory transferred radiology research context between web and local coding environments without copy-paste in a single-user proof-of-concept. The 558-token session had complete six-field extraction; five consecutive GitHub updates had no overwrite errors.

• The pipeline integrates a web app, n8n automation, Claude API, GitHub, and Obsidian; sessions are summarized into six structured fields (summary, decisions, pending tasks, artifacts, errors, continuation point) stored as MEMORY.md and CLAUDE.md files.
• Proof-of-concept is explicitly limited to a single user, single operator, text-only Claude-based use, without quantitative extraction-accuracy benchmarking; commercial API costs scale with use.
• The system has been used in practice to help organize development of a separate radiology report quality-control tool; authors recommend a vendor-neutral, open-weights configuration for cost- and dependency-sensitive settings.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.