The awesome-infra-for-ai list received 16 new repositories. doobidoo/mcp-memory-service (1,901 stars) — Self-hosted persistent memory backend for AI agents that stores decisions and context in a semantic store with a typed knowledge graph, served over MCP, REST, CLI and a dashboard. FailproofAI/failproofai (1,416 stars) — Observability and policy enforcement for AI agent harnesses, hooking twelve coding and chat harnesses to record every run and block dangerous tool calls before they execute. Jwuthri/Tracely-ai (1,048 stars) — Trace-native CI/CD for AI agents that grades production traces, clusters failures, freezes bad runs into hermetic replayable regression cases and blocks the pull request that would ship them again. wangeai/K3Flight (1,019 stars) — Single-file Linux inference server that runs a 2.8 trillion parameter Kimi K3 checkpoint on CPU by streaming weights from storage, keeping the measured resident working set near 55 GB. Continuum-AI-Corp/OrcaRouter-Lite (590 stars) — Self-hosted OpenAI-compatible LLM router that fronts more than a hundred models with bring-your-own-key access, streaming, automatic model selection and an optional hosted fallback. ggozad/haiku.rag (584 stars) — Agentic retrieval-augmented generation library for local document collections, combining hybrid vector and full-text search, reranking and multimodal retrieval on an embedded LanceDB store. ARahim3/mlx-dspark (584 stars) — Native MLX port of the DSpark and DFlash speculative decoding drafters, giving lossless multi-fold faster LLM decoding on Apple Silicon with an OpenAI-compatible serving mode. HarnessRouter/harnessrouter (553 stars) — Self-hosted unified API in front of agent harnesses such as Codex, Claude Code and Hermes, implementing the Unified Harness Protocol with sessions, streaming, files and cancellation. decionis/agent-safe-pipeline (533 stars) — Reference architecture and TypeScript library that routes AI agent actions through an independent authorization boundary, turning proposals into policy verdicts, human approvals and single-use grants. Floe-Labs/floe-guard (516 stars) — In-process spend meter and budget gate for AI agents that hard-stops the next model call, voice turn or paid tool invocation before it crosses a configured cost ceiling. askalf/dario (474 stars) — Local OpenAI- and Anthropic-compatible proxy that exposes a Claude subscription to other AI tools, with session-affinity routing, multi-account pooling and drift detection for long agent runs.