Awareness brings local-first AI agent memory with 96% benchmark retrieval score

A developer has released Awareness, a local-first memory layer for AI coding agents such as Claude Code, Cursor, and Windsurf that stores all data on the user's machine as Markdown files with no cloud account or API key required. The tool automatically injects memory workflows into IDE configurations, so agents retrieve and record context across sessions without manual wiring by the developer. Awareness organises memory into 13 structured categories — including decisions, pitfalls, and workflows — with built-in conflict detection and deduplication. It achieved a 96.0% R@5 score on the LongMemEval benchmark across 500 questions, measured through its own production retrieval pipeline, with a public runner available for independent reproduction. The underlying stack uses a Node.js daemon with SQLite, FTS5 full-text search, and ONNX embeddings, and optionally supports on-chain state anchoring on the Sepolia testnet for verifiability.