Open Source Radar — October 14: memory, small GPUs, big worlds
Half of today's trending list is about memory — what an agent keeps and what it throws away. The other half puts heavy creative and spatial work back on your own machine.
Context Mode (TypeScript, ~28,600 stars) — An MCP server that stops raw tool output from eating your context window: it sandboxes the bulky results behind a query layer and reports a 98% cut, turning a 315 KB dump into roughly 5.4 KB. It also logs every file edit, git operation and decision into a local SQLite database, and when the conversation compacts it retrieves only the relevant events by search instead of dumping them all back in. The third idea is the interesting one: rather than reading fifty files to count something, the agent writes a small script and prints only the answer. It ships hooks for 17 clients, including Claude Code, Codex, Cursor and Zed, plus the OpenClaw gateway — but read the licence before you build on it, because Elastic License 2.0 blocks offering it as a hosted service.
Claude-Mem (TypeScript, ~100,000 stars, Apache-2.0) — Persistent memory for coding agents, built on the same hooks that fire during a normal session: it captures what the agent did, compresses it into semantic summaries, and stores those in SQLite alongside a vector index, so the next session starts already knowing your project. Search is deliberately staged to save tokens — a compact index first, then the timeline around the interesting hits, then full detail only for the entries you actually want, which the project puts at about a tenfold saving over fetching everything. Version 13.39.0 shipped October 13, and at roughly 100,000 stars it is one of the most-starred pieces of agent tooling on the platform. Use it if you are tired of re-explaining your codebase every Monday.
AirLLM (Apache-2.0, ~38,600 stars) — Runs models far larger than your GPU by streaming the weights one layer at a time from disk instead of holding them in VRAM. The README's claims are aggressive: a 70B model on a 4GB card with no quantization, Kimi K3 at 2.8 trillion parameters under 4GB, DeepSeek-V3's 671 billion at around 12GB — and a September update added training, with a 125B model fine-tuned under 6GB. The trade is speed, and it is a real one: nothing here runs fast, so it is the wrong tool when you care about throughput. It is the right tool when the model has to be that big and the card has to be that small.
God's Eye View (JavaScript, ~56,000 stars) — A spy-satellite simulator in a browser tab where the feeds are real: live aircraft, ships, satellites, earthquakes and public cameras on a photorealistic globe, driven hands-free by a realtime AI agent that can count, filter and rank whatever you are looking at by voice. It starts with no API keys at all — imagery, terrain, flights, satellites and quakes work out of the box — and installs in one click through Pinokio or from a terminal on localhost. The documentation is refreshingly honest about the seams: traffic is simulated along real roads and camera positions are estimates, so treat the overlay like any other map. It hit number one on GitHub Trending in August and is still landing commits today.
ArtCraft (Rust, ~26,800 stars) — A desktop "IDE for artists" that builds the scene before you generate it: compose in 2D, stage props and backdrops in 3D, pose characters and place the camera, then hand that locked composition to an image or video model. It fronts a catalogue of 62 models across its own service plus Grok, Midjourney, Sora and World Labs, so you are not married to one vendor. The point is control — setting a character and a camera beats rerolling prompts until a face lands in the right spot. Free to download, and the source is on GitHub if you would rather build it yourself.
Worth watching this week.
Memory layers or smaller footprints, tiny cards or big scenes — which of these are you actually installing? Tell us in the comments.
