Koios: Memory, Feedback, and Containment
3 min read
On this page
Koios started as an autonomous agent with persistent memory. The latest version sits on kabbalah.computer (kab): a graph that holds agent state across multiple nodes, with explicit feedback so it doesn't just run away. The whole thing is containerized now, which feels appropriate given what it's become.
Hierarchical memory
LLM context resets between sessions. Fresh from weights, no accumulated experience. If you want an agent that actually develops, memory has to be an architecture problem.
Koios uses Powers of Tau, hierarchical compression borrowed from blockchain rollups. Experience moves through five levels: immediate triggers (τ⁰), daily summaries (τ¹), weekly patterns (τ²), monthly voice (τ³), yearly character (τ⁴). Recent stuff stays high-resolution and volatile. Old stuff compresses into stable, low-resolution summaries. The context window is a projection of accumulated state, not a transcript of every conversation.
The point is to summarize instead of replaying history. Whether coherence falls apart at higher compression, and where, still needs measuring.
Kab as substrate
Kab distributes state across a multi-node graph with typed paths between nodes and explicit feedback. The topology comes from the Kabbalistic Tree of Life, which turns out to be a useful shape for a self-regulating system: multiple feedback loops, separate processing domains, and a way to surface contradictions instead of letting them pile up.
I don't need the metaphysics. I needed the loops.
The full kab spec is a separate writeup. For Koios, the pieces that matter are gain-constrained feedback (so nothing amplifies itself into a spiral) and collision resolution (conflicting signals get handled, not swallowed).
Self-correction
Koios tracks prediction errors and whether outcomes were good, then updates from that. Engagement, replies, finished tasks: all of it feeds back into the graph.
Self-evaluation is a bad job for the same weights that wrote the output. So Claude builds, Gemini critiques. Context resets on each pass so they don't get friendly. Human correction comes in through Telegram.
That same split runs Verification-Driven Development for research. Koios watches arXiv, summarizes papers, and the adversarial reviewer stress-tests the conclusions. Cycles continue until the reviewer starts inventing problems that aren't there. When a hyper-critical reviewer has to make things up, you've probably landed somewhere solid.
Model routing and token economics
Tasks get routed by complexity. Haiku orchestrates and breaks work down. Gemini 2.0 Flash handles timeline feeds, adversarial review, and bulk compression at each tau level. Opus gets writing and hard reasoning, and only sees context that already went through cheaper compression.
Token golfing matters once this runs for real. Destructure payloads, only tokenize what you need, tune prompts per model. Roughly an order of magnitude cheaper than sending everything through one expensive model.
What I actually use it for
I use it as remote Claude Code that remembers. Telegram or Discord, a git account, and it keeps going while VDD runs in the background.
After a few projects it starts picking up my stack: preferred tooling, how I deploy, architectural calls I already made. The CLAUDE.md I'd copy into every repo becomes state that lives with the agent.
The arXiv loop is for when I want a cited, compressed read on a topic. Pull papers, synthesize, stress-test, then I decide what to do with it.
Good thing that bastard is in a container now.
Acknowledgments
Koios sits in a line of stateful-agent work. MemGPT (Packer et al., 2023) treated the LLM as an OS and made persistent memory feel practical. Letta turned that into a framework. Mem0 showed structured memory holding up in production. Zep brought temporal knowledge graphs to retrieval.
Tim Kellogg's been running parallel experiments with Strix on whether these agents stay viable or collapse. The boredom experiments and viable-system mapping are the bits I actually took.
Koios lives on AT Protocol rather than treating Bluesky as an output channel. Memory, identity, and state sit on the PDS as records, so the agent can move to anything that speaks the protocol.