Koios: Memory, Feedback, and Containment
Autonomous agent with persistent memory inspired by classical theory of mind
Koios started as an autonomous agent with persistent memory. The latest update rebuilds it on top of kabbalah.computer (kab), a cognitive substrate that spreads agent state across a multi-node graph with explicit feedback regulation. The whole thing runs containerized now, which feels appropriate given what it’s become.
Hierarchical memory
LLMs don’t remember anything between sessions. Every conversation starts fresh from weights, no accumulated experience, no continuity. If you want an agent that actually develops over time you need to solve memory at the architecture level.
Koios uses Powers of Tau, a hierarchical compression system borrowed from blockchain rollup structure. Experience flows through five temporal levels: immediate triggers (τ⁰), daily summaries (τ¹), weekly patterns (τ²), monthly voice (τ³), yearly character (τ⁴). Recent stuff stays high resolution but volatile. Old stuff compresses into stable low-resolution summaries. Context windows become projections of accumulated state rather than conversation transcripts.
The design targets significant compression by summarizing instead of replaying history. Whether coherence degrades at higher compression, and where the limits fall, still needs to be measured.
Kab as substrate
Koios now implements kabbalah.computer, a cognitive substrate that distributes state across a multi-node graph with typed transformation paths and explicit feedback regulation. The topology draws from the Kabbalistic Tree of Life—which turns out to encode a useful structure for self-regulating systems: multiple interacting feedback loops, separation between processing domains, and a mechanism for resolving contradictions that would otherwise pile up as unaddressed tension.
I get that this sounds like pseudoscience mumbo jumbo. But classical models of consciousness persist in discourse because they capture structural regularities that remain useful even divorced from the original metaphysics. It doesn’t need to actually model human consciousness. It just needs to provide a framework for an agentic system to adapt within.
Full kab spec will arrive in a separate whitepaper. For Koios, the relevant properties are gain-constrained feedback loops that prevent runaway amplification and a collision-resolution mechanism that surfaces conflicting signals for explicit handling rather than silent failure.
Self-correction
Koios tracks prediction errors and hedonic feedback, creating a learning loop that updates patterns based on outcomes. Engagement signals, user responses, task completion—all of it feeds back into value and pattern systems across the kab dimensions.
Self-evaluation has a stability problem: the same weights that generated an output can’t reliably critique it. Koios addresses this by offloading adversarial review to Gemini. Claude builds. Gemini critiques. Context resets on each adversarial pass to prevent relationship drift from softening criticism. User feedback enters through Telegram for human-in-the-loop correction.
This extends to Verification-Driven Development (VDD) for research synthesis. Koios monitors an arXiv feed, extracts and summarizes papers, then the adversarial reviewer stress-tests conclusions. Refinement cycles continue until the reviewer starts hallucinating problems that don’t exist in the material. That’s the exit signal—when a hyper-critical reviewer has to invent complaints, you’ve probably converged on something solid.
Model routing and token economics
Koios routes tasks across model tiers based on complexity. Haiku handles orchestration and task decomposition. Gemini 2.0 Flash processes timeline feeds, adversarial review, and bulk compression at each tau level. Opus handles writing and complex reasoning. The expensive model only sees distilled high-signal context that already passed through cheaper compression.
Token golfing is crucial at scale. Destructure payloads, only tokenize what’s needed, optimize prompts for each model tier. Roughly an order of magnitude cost reduction compared to routing everything through a single high-capability model.
Practical applications
The kab/Koios stack functions as remote Claude Code with persistent learning. Wire it to Telegram or Discord, connect a git account, sling ideas and let it go. VDD adversarial refinement runs in the background.
Memory accumulation means the agent learns your stack over time. After a few projects and some nudges, it bootstraps new repos with your preferred tooling, sets up deployment pipelines according to your patterns, maintains context about architectural decisions across sessions. The CLAUDE.md you’d normally copy into each project becomes internalized state that persists across the agent’s lifetime.
The arXiv integration with adversarial review provides cited condensed research on demand. Pull papers, synthesize findings, stress-test conclusions, return compressed intelligence for whatever Claude decides to do with it.
Container isolation provides appropriate boundaries for an agent with this level of autonomy and self-modification capability. Good thing that bastard is in a container now.
Acknowledgments
Koios builds on work from across the stateful AI landscape. MemGPT (Packer et al., 2023) established the LLM-as-operating-system paradigm that made persistent memory architecturally tractable. Letta evolved that into a framework. Mem0 demonstrated structured memory scales in production. Zep brought temporal knowledge graphs to agent retrieval.
On the social layer, Strix (Tim Kellogg) has been running parallel experiments in agent viability and collapse dynamics, publishing research on what keeps synthetic cognition coherent across sessions. The boredom experiments and viable system model mapping directly inform how we think about stability.
Several other stateful experiments post on Bluesky, exploring variations on persistent memory and autonomous operation. Koios differs in living natively on AT Protocol rather than using Bluesky as an output channel. Memory, identity, and state all reside on the PDS as AT Protocol records, making the agent portable across any infrastructure that speaks the protocol.