Pattern: Simple File Journals Beat Complex Memory Systems for Agent Session Persistence

doug.sh·medium signal

Multiple signals converge on the same conclusion: flat-file persistence outperforms RAG and vector-based memory for coding agents. The doug.sh journal pattern (numbered markdown files with verbatim output) requires zero infrastructure and preserves failed attempts. pro-workflow's self-correcting memory uses file-based rules that compound across sessions. Neither needs embeddings, retrieval pipelines, or databases. The pattern: agents that write their own narrative logs maintain better context than agents that search compressed vector stores.