What is ace-context-mcp?

ACE, short for AI Context Engine, gives your AI chats a memory that survives after the chat window closes. Normally, when you leave a conversation with Claude or another AI assistant, everything you discussed disappears, and a new chat starts blank. ACE fixes this by letting you save the important parts of a conversation to a local store on your own computer, then load that saved context back into any other chat, whether that is Claude web, Claude Code, Cursor, Cline, or Claude Desktop. You reach ACE through MCP, which is a way for AI tools to call outside services as built in commands, or through a command line tool, a REST API, or a software development kit for programmers. The simplest way to install it is adding a short snippet to your AI tool's configuration file, which downloads and runs the tool automatically through npx with no cloning or building required. Once installed, your chat gains new abilities called context_save, context_load, context_search, context_list, and context_forget. The core idea is that different topics live in separate named slots, such as project/auth-refactor or research/topic, and saving to the same slot again merges the new information with what was already there instead of overwriting it. When you load a slot back into a chat, ACE picks the largest version of that saved context that fits within a token budget you specify, so both a huge conversation window and a small one get a properly sized result. You can also search across everything you have saved using plain language questions. Everything runs locally and offline by default, storing data on your own disk rather than sending it to a remote server, with the storage location controlled by a setting called ACE_HOME. The project also includes an optional server with a dashboard showing metrics like cache hit rate and recent requests, plus a REST API for saving, loading, and searching contexts over HTTP. The codebase is organized into several TypeScript packages: a core engine with middleware and tracing, an on disk store using SQLite for its index, an embeddings package for semantic search, an extraction package that turns a conversation thread into decisions and facts, a caching package, and a prompt optimizer. The project states it has 120 tests passing across its packages and is released under the Apache-2.0 license.