Open Source Radar — September 25: memory, method, smaller models
Today's trending list splits three ways: what an agent remembers, how it works, and whether any of it fits on hardware you already own.
Hindsight (Python, ~28,400 stars, MIT) — Vectorize's memory layer is one of the most-trended repos on GitHub today, and the distinction it draws is the useful part: most agent memory is conversation history, while this is built so an agent's beliefs change as evidence accumulates. Facts go in through one call, get consolidated in the background into observations that keep their supporting quotes and a proof count, and are refined rather than overwritten when something contradicts them. Recall runs vector search, keyword matching, graph traversal and time filtering in parallel, then merges and reranks the results, and a second operation synthesizes conclusions that need thinking rather than lookup. The practical payoff is the boring feature: a "mental model" is a standing answer to a question like what this user's preferences are, written in the background and read back as a plain database read, so an agent boots with settled knowledge instead of rediscovering it every session. It runs embedded in Python, in Docker or Kubernetes, or on the hosted tier, and it will talk to a local model through Ollama or llama.cpp as readily as to a frontier API — subscriptions you already pay for count too. The caveat: the benchmark lead is the vendor's claim, and the scores in its comparison chart are self-reported, though the team says Virginia Tech's Sanghani Center and The Washington Post reproduced its long-term-memory numbers independently.
Superpowers (Shell, ~291,000 stars, MIT) — The largest number on today's trending page belongs to a skills framework, not a model, and it is the clearest statement of what the agent-skills idea is actually for. Instead of letting a coding agent start writing the moment it sees a prompt, it makes the agent pull a spec out of the conversation first, show it in chunks short enough to read, then produce an implementation plan written for an enthusiastic junior with no project context — and only then start building. From there the work runs as a sequence of subagents with two-stage review, spec compliance before code quality, which is how the README can claim agents running for a couple of hours without drifting from the plan. It installs into a long list of harnesses — Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Qwen Code, Kimi Code among them — and the skills fire on their own rather than waiting to be invoked. Two things worth knowing before you standardise on it: there is a commercial support arm behind the project, and an optional visual feature reports the version in use back to the authors, with an opt-out documented.
stable-diffusion.cpp (C++, ~7,300 stars, MIT) — The image-generation counterpart to llama.cpp, and what puts it back on the list is the pace of the model list rather than any single release: Qwen-Image-2.1 support arrived on day zero on September 20, LTX-2.5 in August, MiniMax-H3 a day after it appeared. The whole thing is plain C and C++ on ggml with no external dependencies, which is what lets it run on a laptop, a Windows machine, an Apple Silicon Mac, or an Android phone through Termux — and it covers far more than Stable Diffusion now, including FLUX.2, Qwen-Image and its edit variants, Z-Image, Ideogram4 and Krea2, plus video models such as Wan2.1 and 2.2 and HunyuanVideo 1.5. The memory controls are why people stay: attention and tiling techniques to squeeze large models onto small cards, GGUF quantization, LoRA, and an embedded web interface so you are not writing Python to test a prompt. It has grown its own ecosystem too — bindings for Go, Rust, C#, Python and Flutter, and generation backends used by LocalAI and KoboldCpp. The warning is printed in the README: active development, and the options change often.
NVIDIA Model-Optimizer (Python, ~4,200 stars, Apache-2.0) — Quantization, pruning, distillation, sparsity and speculative decoding in one library, and version 0.47.0 shipped two days ago. The reason to care is what the published recipes measure: an end-to-end tutorial for a 35-billion-parameter mixture-of-experts model at 4-bit weights and activations reports about 1.3x the throughput of the full-precision version in vLLM with checkpoints roughly a third the size, and NVIDIA's own 550-billion-parameter Nemotron 3 Ultra checkpoint claims up to 5.9x higher decode-heavy throughput than a comparable FP4 model while matching its full-precision accuracy. Compression only counts if it deploys, so the output targets vLLM, SGLang, TensorRT-LLM and TensorRT instead of staying in a notebook, and a collection of ready-made quantized checkpoints is published on Hugging Face for anyone who would rather download than run the recipe. It is the least glamorous useful project in open source right now, and it answers the question every team eventually asks: can a bigger model run on the hardware we already have. Caveat from the repo's own policy — still pre-1.0, breaking changes allowed in minor releases, and about a month of notice before deprecated features disappear.
Worth watching this week: whether Hindsight's memory lead survives independent testing, and which new image model gets day-zero support next.
Does better memory or better methodology buy your agents more — and which one are you actually missing? Tell us in the comments.
Sources: Hindsight (GitHub) · Hindsight paper (arXiv) · Superpowers (GitHub) · stable-diffusion.cpp (GitHub) · NVIDIA Model-Optimizer (GitHub) · NVIDIA Model Optimizer docs
