thr: Local semantic memory for coding agents

The landscape of AI development, particularly involving agent workflows, is increasingly reliant on context. While large Language Models (LLMs) provide remarkable reasoning power, they often lack persistent, personalized memory of a project or a developer’s unique history. Enter `thr`, a lightweight command-line interface designed precisely to address this knowledge gap. At its core, `thr` functions as a local, structured memory bank for the terminal—a system to save explicit memories that can be recalled later either by precise text search or, more powerfully, by semantic similarity. What sets `thr` apart in a crowded market of developer tools is its unwavering commitment to a local-first architecture. Unlike tools that might push all memory operations through a cloud endpoint, `thr` stores its data (memories, embeddings, metadata) within a local SQLite database. This design makes it inherently private and reliable, crucial for developers handling sensitive project facts. The core functionality is driven by a sophisticated embedding model bundle, enabling robust offline semantic search—a feature that allows a user to ask, 'What was the design constraint on the networking module?' and retrieve memories based on *meaning*, even if the exact query terms were never used when the memory was saved. Functionally, `thr` is an engine for agent augmentation. Its stable JSON output and the provided `thr setup ` commands are not mere features; they are enabling layers. They allow external coding agents (like those powering advanced IDE features) to safely and predictably interface with the developer's accumulated knowledge. Commands like `thr ask --json` or `thr search --json` guarantee structured data that an agent can ingest and use for informed decision-making, making it a true developer utility rather than just a note-taking app. While the CLI commands are straightforward—`thr add`, `thr ask`, `thr forget`—the technical implementation depth is what warrants attention. The inclusion of bundled runtime assets (like ONNX Runtime and the embedding model) ensures that the sophisticated search capabilities remain functional even when the user is completely disconnected from the internet. This local capability is a major selling point for enterprise and security-conscious development teams. For the developer, `thr` is not just a place to store notes; it's an infrastructural component for maintaining project continuity and enhancing agent reliability.