Developer Builds Git-Committed AI Coding Memory System After 3 Months, 542 Lessons

A developer who identified a core flaw in AI coding agents — their inability to retain project-specific corrections between sessions — built a lightweight graph-based memory system to address it. The system stores single-sentence lessons tied to specific files, commands, or keywords, and surfaces them to the agent at the exact moment they are relevant, not at session start. After three months, the repository holds 542 lessons, most written by agents themselves, with over 5,000 recalls logged. Recent improvements include specificity-ranked recall, whole-word keyword matching, and semantic understanding of task intent beyond literal keywords. The tool is open source under the MIT license, requires no vector database or network calls, and costs roughly ten tokens per recall.