NMEMORY: An AI Agent Memory That Refuses to Guess

NMEMORY is a single-file SQLite memory store that AI agents talk to over MCP (stdio), built to solve the recurring problem of agents losing all context between sessions. Every captured fact is stored with its source and anchor, and recall never returns more than the evidence supports: the system defines exactly three honest outcomes — grounded (matched real records), missing_evidence (matched but excluded), and abstain (no match) — with no fourth 'best guess' option.

The design treats trust as the primary feature rather than recall volume. Retrieved memories are always wrapped as advisory data, never rendered as instructions, so stored content cannot hijack the agent. The serving path is compiled with zero networking code, runs no daemon, and keeps data in one local SQLite file the user fully owns; syncing across machines is an explicit, owner-triggered action rather than a background process.

Written in Rust and backed by a 589-test hermetic suite, the project publishes concrete verification steps (strace/ldd checks for zero-network, dedicated tests for zero-Python and deterministic exports) rather than asking users to trust claims blindly. It exposes 21 MCP tools across four planes — capture, recall, structure, and lifecycle — and supports remote setups over SSH while keeping the store itself single-host. For engineers running agents in production, it's a concrete example of a memory layer designed to say 'I don't know' instead of fabricating an answer.