The fastest-growing open-source repo in AI right now is a memory plugin. claude-mem hit ~50K GitHub stars because it solved a problem Anthropic hadn't shipped yet: persistent context across Claude Code sessions. One install, and every decision you made, every bug you fixed, every architectural choice from Tuesday's session is waiting for you Wednesday morning. The "95% less tokens" claim is real but requires context. Claude Code's context window is 200K tokens (1M on Opus). When you hit ~95% capacity, the system auto-compacts by summarizing your history. That summary is lossy. Instructions you established early vanish. Corrections disappear. Patterns you spent 30 turns teaching get compressed into "user prefers certain coding conventions." claude-mem sidesteps the entire compaction problem. It runs a background observer that captures tool usage in real time, generates semantic summaries, stores them in a local SQLite database and Chroma vector store, then injects only the relevant observations back at session start. Your session starts with a lightweight index. The LLM fetches full records only when it needs depth. The architecture is the interesting part. Five lifecycle hooks (SessionStart, UserPromptSubmit, PostToolUse, Summary, SessionEnd), a worker service on port 37777, and a skill system that lets you search past work with natural language. The creator also launched a Solana token ($CMEM) alongside it, which tells you exactly where the open-source incentive model is heading in 2026. 49,500 stars in roughly two weeks. 3,900 forks. 1,662 commits. 117 pull requests. Alex Newman built the most popular Claude Code plugin in existence because every developer who's ever lost context to auto-compaction felt the same pain, and nobody upstream had fixed it yet. The fastest path to a massive open-source project right now is finding the gap between what an AI company ships and what developers actually need in production. Memory was that gap.