🔥 Hot Repo: 67K Stars — Claude Finally Remembers Between Sessions

If you have ever restarted a Claude Code session and spent the first ten minutes re-explaining your project architecture, thedotmack/claude-mem is the 67,000-star plugin that eliminates that tax. The stateless problem Claude Code is stateless by design. Each new session begins cold — no memory of last week's debugging session, no recall of the architectural decision that ruled out option B, no context about the half-finished refactor. For developers working on long-running projects, this means repeatedly rebuilding context that the agent should already have. How claude-mem works The plugin hooks into Claude Code's six lifecycle events: SessionStart, UserPromptSubmit, PostToolUse, Stop, and SessionEnd. At each hook point, it records typed observations — bugfixes, decisions, errors, security notes — and uses the Claude agent-sdk to compress them into semantic summaries stored in a local SQLite database with full-text search FTS5. At the start of a new session, relevant past observations are retrieved and injected into context automatically. No manual intervention, no re-explaining. The backend is a Bun-managed worker service running on localhost:37777 with a real-time web viewer. Alongside text search, a Chroma vector database powers hybrid semantic + keyword retrieval. The 4 MCP tools follow a three-layer retrieval pattern: 1. search — returns a compact index 50–100 tokens per result 2. timeline — chronological context around specific hits 3. getobservations — full details for filtered IDs only The pattern delivers roughly 10x token savings compared to dumping raw history into the prompt. Installation is one command: bash npx claude-mem install Or directly inside Claude Code: bash /plugin marketplace add thedotmack/claude-mem Support also covers Gemini CLI and OpenCode. What v12.3.9 changed The release landed April 22 and addressed two high-friction issues. The most impactful fix: a 110-second terminal block that occurred every time a Claude Code