menot-you/n-memory

nMEMORY is a memory tool designed for AI coding assistants. The core problem it addresses is that AI agents typically start each session with no recollection of prior decisions, design choices, or past bugs, forcing the developer to re-explain context every time. This project, written in Rust, gives an agent a persistent, local memory store that it can consult, with a single strict rule: when the store does not have an answer, it says so plainly rather than guessing. The tool is built around the idea that a memory which returns a plausible-sounding but unverified answer is worse than no memory at all. When an agent stores information, it must include the origin and a reference point for that data. Any memory entry lacking a source is rejected. When an agent later retrieves information, the system returns one of three outcomes. It either provides the matching evidence with its source and freshness attached, reports that matches were found but excluded (for example, if the information was superseded), or it abstains entirely and states it has nothing relevant. The system is designed to operate entirely on your local machine. It uses a single SQLite file to store all data, with no server, no account, and no background processes running. The binary is compiled without a networking stack, meaning it makes zero socket connections and sends no telemetry anywhere. If you want to move your memory file to another machine, you must do it yourself using an explicit sync command, which delegates the file copy to a separate process. You install it by running a shell script or building it from source with Rust. Once installed, you register it with your AI agent as an MCP tool. The agent then has access to 21 different tools for storing, retrieving, and managing memory. Key operations include capturing new memories with their sources, recalling past information as evidence, and viewing a session summary of what the system knows, what is ready, and what is blocked. Everything the memory returns is labeled as advisory data, not as an instruction, so the stored text cannot hijack the agent's behavior. The full README is longer than what was shown.