ContextKeep solves the "AI with amnesia" problem by giving Claude, Cursor, and other MCP clients a persistent local memory layer. It exposes eight tools over stdio or SSE: list_all_memories, retrieve_memory, store_memory, search_memories, and newer additions like delete_memory and export_memories. The two-step retrieval pattern (list keys first, then fetch by exact key) prevents the common failure mode where agents hallucinate memory IDs. You get a web dashboard on port 5000 for visual management, Docker Compose support for homelab deployments, and everything stores locally as JSON. Reach for this when you're tired of re-explaining project context every conversation or want to slash token costs by pulling only relevant memories into each request.
claude mcp add --transport stdio mordang7-contextkeep uvx contextkeep