Launching today
Auto Learning Agents
Self-hosted AI agent OS with memory that actually learns
7 followers
Self-hosted AI agent OS with memory that actually learns
7 followers
Auto Learning Agents is a free, open source OS for AI agents. Deep internal memory plus dynamic, session-less context: agents find the right context each turn, pick up where they left off, and get better at recurring work. Web UI for Claude, OpenAI, Gemini, or fully local with Ollama, with a file browser, code editor, and 40+ built-in tools. Background agents run while you are away. Built on Elixir/OTP, every agent is a supervised process that recovers from crashes. One Docker command and go.
How does the deep internal memory actually work across sessions, do agents truly retain learnings between runs or is it more like scoped context retrieval each turn?
@hiraxevj Well several methods, but scoped retrieval is the main part. So for memory bank, it remembers things you tell it to, general_knowledge, user_knowledge, callable_scripts, etc which is a local embeddings server with a tool claude can call to look things up. Then there is conversation memory, which lets you create conversation threads, it saves and embeds every turn. The local embeddings is the key though, not about the cost, local makes the queries SOOOO fast, that claude falls in love with the tool and calls it often.
So conversations claude/codex looks up by itself, your settings determine how many of the last messages to send automatically, like 5, and then if the agent needs more it can either search all context or just pull a few more previous turns.
Memory is such a generic word that it's tough to keep using that word, but there's also memory files which is a standard thing claude code does on its own, this software just organizes those for you. For context the goal is for claude to have access to all of it, but not just call it on its own or it will cross reference things you don't want it to, so memory bank is generally you telling it "look up this ..." so it doesn't wildly know everything. Personally I have like 50+ projects in mine so I have my own rules in claude.md that says "ask me before pulling anything" cause my projects have things I don't want mixed, but if you set it up for say your business only, you can tell the AI in the rules what it can look up on its own.
The OTP supervision per agent is such a thoughtful call. Most agent frameworks just hope nothing crashes mid-run, but treating each one as a supervised process that actually recovers feels like the kind of foundational decision that will make this hold up once people build real things on it.
@kazmeriifqw Thanks Kazım, ya the genserver foundation once I learned it, is now in anything I write that's multiple processes. Also, the interactive genserver (tools/interactive.php) lets claude run interactive commands, so claude can run a multi command line input session because the genserver "holds it" in between calls. It has another one too for ssh using python, you can make claude manage an entire remote server with it.
Some of you beat me to first comment haha, thank you for the support I have already received, its only been 3 hours and I got about 50 messages I woke up to on various social media so catching up now.
I built Auto Learning Agents because my agents kept forgetting everything between sessions. It is a self-hosted agent OS built on Elixir/OTP: deep internal memory, session-less dynamic context so agents find the right context each turn, a web UI for Claude, OpenAI, Gemini or fully local Ollama, plus a file browser, code editor, 40+ built-in tools, and background agents that keep working while you sleep. Free and open source, one Docker command to run. I would love honest feedback, especially from anyone self-hosting their AI stack, and I will be around all day answering everything.