Hi everyone — maker here. I built Veles because of a gap that kept bugging me: AI tools are great in the moment, but the understanding never accumulates where it matters — in the project itself. Most "memory" today is chat-scoped, cloud-locked, and opaque: you can't see it, edit it, or take it with you. So every serious session still starts with me re-explaining the same context. Veles makes memory the core feature, not an add-on. After each session a "Curator" distills what happened into structured per-project memory you can actually read and edit — plain files in `.veles/`: insights, behavioral rules, decisions, session digests. Next time it recalls the relevant ones on its own. The more you work on a project, the less you repeat yourself. A few principles I held to: • Memory-first — the learning loop IS the product, not a plugin. • Yours, locally — no cloud, no sign-up; everything lives in your repo. Run it fully local with Ollama/llama.cpp, or plug in OpenRouter/Anthropic/OpenAI/Gemini — even your existing claude/gemini CLI subscriptions. • Clean code — no 12k-line god-files. Apache-2.0, easy to read and extend. • Pluggable — a Karpathy-style LLM-wiki layout by default, flat notes, or nothing at all for code repos. Optional TUI, daemon + Telegram, multi-project. Try it (Python 3.13+): uv tool install veles-ai veles init && veles It's free and fully open-source. I'll be here all day answering everything. One thing I'd genuinely love your take on: how much should an agent be allowed to remember and act on automatically before it asks you first? Where's the line between helpful and creepy? ⭐ https://github.com/denisotree/veles

Veles

Open-source CLI agent with memory you own and can read

Open-source CLI agent with memory you own and can read