🔥 Hot Repo: 57K Stars — The Plugin Curing Claude Code's Amnesia

Every developer who uses Claude Code knows the frustration: you spend an hour debugging a gnarly auth issue, Claude fixes it perfectly — and then your next session starts completely blank. No memory of what was decided, what was tried, what patterns the codebase follows. You are back to square one, re-explaining your entire project to an AI that should know it already. claude-mem is the open-source fix that has earned 56,937 GitHub stars since its August 2025 debut — and it added roughly 10,000 of those in the past seven days alone. A new maintenance release, v12.1.3, dropped today April 15, 2026, as the project continues to mature rapidly. How It Works claude-mem installs as a Claude Code plugin in a single command: bash npx claude-mem install Or directly from within Claude Code: /plugin marketplace add thedotmack/claude-mem Once installed, five lifecycle hooks SessionStart, UserPromptSubmit, PostToolUse, Stop, SessionEnd silently intercept every tool call Claude makes. Observations are streamed to a local worker service running on port 37777 powered by Bun, stored in SQLite with FTS5 full-text search, and indexed in a Chroma vector database for semantic retrieval. At the start of each new session, the most relevant past observations are fetched and injected into Claude's context automatically. The Search Architecture The MCP-based search layer follows a 3-step, token-efficient workflow that the project claims delivers roughly 10x token savings over naive full-context injection: 1. search — returns a compact index 50–100 tokens/result 2. timeline — retrieves chronological context around results of interest 3. getobservations — fetches full details only for the observations you actually need 500–1,000 tokens/result This means Claude can query its memory efficiently without blowing up the context window on irrelevant history. What v12.1.3 Reverted Today's release is a targeted rollback: it removes "overengineered summary salvage logic" added in PR