Open Source Radar — August 21: memory, vectors & red teams
The daily board is betting on the layer under the agents again — where they remember, how fast they search, and whether they're safe to point at your stack. Five fresh open-source projects worth a look.
OpenViking (Python, ~31.3k stars) — Volcengine's context database for agents that tries to be the one place memory, retrieval, and skills live together, and it's built to evolve itself as it runs. The pitch is that today's agents relearn the same context every session because memory, RAG, and tool-use sit in three different places, and OpenViking collapses them into a single self-updating store. You'd drop it in as the persistent brain behind a coding or research agent so it stops rediscovering what it already knew. If agent memory has felt like three half-built features bolted on, this is the consolidation play.
turbovec (Rust, ~16.1k stars) — A vector index built on top of TurboQuant with Python bindings, pitched as a faster, leaner answer to FAISS for nearest-neighbor search. The interesting part is the quantization layer: it shrinks embeddings so similarity search stays quick and light on memory, which matters the moment your RAG corpus outgrows what fits in RAM. You'd reach for it when your retrieval step is the bottleneck and you don't want to stand up a full vector database. It's early, but the star velocity says the search-speed crowd is paying attention.
AI-Infra-Guard (Python, ~5.2k stars) — Tencent's open-source red-teaming platform for AI systems, scanning agents, MCP servers, and skills for the vulnerabilities that actually bite — prompt injection, jailbreaks, and misconfigured tool access. It covers agent scan, skills scan, MCP scan, and an LLM jailbreak evaluation, which is a tidy map of where the risk concentrates as agents get wired into real infrastructure. You'd run it the way you'd run a dependency scanner, but pointed at your agent stack instead of your lockfile. With MCP adoption outrunning its security tooling, this is the audit layer the ecosystem has been missing.
munder-difflin (TypeScript, ~3.3k stars) — A local, free multi-agent harness that runs several agents over the same work with shared memory, built for people who want orchestration without a hosted control plane. The angle is local-first: your agents coordinate on your machine, keep a memory of what they've done, and don't ship your tasks to someone else's server. You'd use it to chain agents on a project — research, draft, review — without writing the glue yourself. It's a small repo but a clear bet on owning your agent runtime instead of renting it.
substrate (Go, ~1.5k stars) — The core system of the Agent Substrate project, written in Go — a foundation for building agentic applications that isn't tied to the Python stack most harnesses assume. The repo's own description is deliberately spare ("the core system"), so treat this as early and watch what gets built on top rather than what ships today. The play, if it lands, is a Go-native base layer for agents where the other picks on this board are Python or TypeScript. Worth a star if you want agent infrastructure that runs where your backend already does.
Worth watching this week: the board is rewarding agent plumbing — memory, search, security, orchestration — over another chat wrapper.
If you had to harden just one layer of your agent stack this week — memory, search, or security — which wins? Tell us in the comments.
Sources: OpenViking (GitHub) · turbovec (GitHub) · AI-Infra-Guard (GitHub) · munder-difflin (GitHub) · substrate (GitHub)
