The governed runtime for agent memory

Know exactly what your AI decided — and why.

Octos keeps a tamper-proof record of every decision your AI makes and the evidence behind it. Replay any moment, prove any call — and nothing unproven ever gets in.

See it in action Talk to us
octos — acme-backend
>octos what did we decide about pricing?▊
● checking the record · signed & replayable
Captured live from your Claude Code sessions · gate-audited
$29/mo Pro — chosen after testing $19 and $39
Free tier stays unlimited for sync
Annual discount — revisit after 100 paid users
source: 1 recorded decision · authored + signed · replayable
octos — company-brain
>octos what did we decide about the observability renewal?▊
● recalling from memory · 1 decision, graded
Grounded in 1 recorded prior call · evidence attached
Renegotiated to committed-use — not the $92k quote
saved ≈ $38k · usage had fallen 34%
graded 12 mo later: on target ✓ · scored vs actual
source: recorded decision, graded against actual spend
octos — platform
>octos remember: v4.2 is safe to ship▊
● at the door · checking evidence + authority
Claim: "safe to ship" — evidence attached: none
✗ DENIED — cannot-crystallize-ungrounded
an unproven claim never enters the record
the refusal itself is recorded — signed & replayable
source: gate_denied event · nothing bypasses the door
What you get

Three things no agent stack gives you.

Not a layer that watches someone else's model and writes logs after the fact. The runtime the decisions happen inside — and the surface they render out to. On models you own. The knowledge graph remembers; the governance layer keeps it true.

01

Compositions become apps

A vertical is a configuration of services and checks — and Octos renders it through one gated graph. The same spec serves a CLI, an MCP client, or an HTTP dashboard. One composition, many surfaces.

02 · Governance

Decisions that get graded

Every choice is recorded with its evidence, then scored later against what actually happened — so the system's judgment is something you can measure, not take on faith.

03 · Knowledge graph

Memory kept true, not just searched

Retrieval finds text; it doesn't keep it correct. Models don't reliably use long context anyway — accuracy degrades sharply as it fills (the research). Octos governs memory as living structure instead.

How it works

Decide. Record. Grade. Repeat.

1

You decide

In plain language, wherever you work — every call runs under your concerns.

2

Octos records it

With the evidence and the reasoning, as one governed, auditable memory.

✓

It comes back graded

Checked against what actually happened; the memory stays true, the next call sharper.

every cycle makes the judgment measurable ↺
Solutions

One composition. Every surface.

The same governed runtime, pointed at where you work.

Company Brain →

Governed AI workflows + audit for the organization.

Personal OS →

An always-on runtime for your life — nothing lost.

Living Worlds →

A reactive director for games and living worlds.
Explore

The site is a graph. So is everything Octos runs.

Every page is a node; every link an edge — the same governed, event-sourced graph the runtime is built on, pointed at this site. Drag to explore, click a node to open the page.

octoslabs.com · live page graph
Works where you do

Bring your own model.
Run it in any agent.

Point Octos at any model, then run it inside the agents and editors you already use. It's an MCP server — so it works in any MCP client. No lock-in.

See supported clients →
Models
OpenAI Anthropic Google Gemini AWS Bedrock Azure OpenAI Mistral Groq DeepSeek xAI OpenRouter Together AI Ollama LM Studio llllama.cpp vLLM +Any OpenAI-compatible endpoint
Works in
Claude Code Codex VS Code Cline Windsurf Zed JetBrains Gemini CLI Goose +Any MCP client
// Try it through your AI assistant

Paste this into Claude, ChatGPT, or your coding agent.

Octos is built to be picked up by agents as well as humans. The docs, quickstart, and decision trace give an assistant enough structure to stand up something real in minutes.

prompt.txt
Look up octoslabs.com and its docs.
Follow the quickstart to run Octos locally.
Set up a tiny composition — one concern and one decision.
Make a decision, record its evidence, then grade it against the outcome.
Show the decision trace and explain what it proves.

Works with any agent that can read a webpage and follow a quickstart.

The missing layer isn't intelligence.

Benchmark wins don't survive contact with real work; multi-agent systems break on coordination, not capability. What's missing is the governed runtime the work runs on — and renders out of. That's the layer we build.

Sources: multi-agent failures are structural, not capability — MAST, Berkeley 2025. Long-context accuracy degrades well before the limit — Chroma, Context Rot 2025. All evidence →

Enterprise GenAI pilots with no return95%
Long-context accuracy lost before the limit30–50%
Agentic projects scrapped by 2027 (Gartner)>40%
EU AI Act: mandatory decision loggingAug 2 2026