SIBYL LABSdocs
Sibyl Labs · Research and Engineering

Building infrastructure for AI agents. and the humans who work with them.

Sibyl Labs is a research and engineering company. We build three things: the memory that lets AI agents remember you between conversations, the substrate that gives them a real identity instead of a blank slate every time, and the tools that turn AI from a chat box into something that actually amplifies what you can do.

Founded 2026 · Local-first by design #2 globally on LongMemEval Oracle, paired with Claude Opus 4.6

The three things AI agents are missing today

When you try to use an AI agent the way you'd use a real collaborator. Not as a one-shot chatbot, but as someone who shows up, knows your context, has opinions, and gets sharper the longer you work together. You run into three walls.

01 Your agent forgets you.

Most agents start every conversation from zero. The ones that don't forget store everything on someone else's servers, recall it as one fuzzy blob, and quietly lose the important context when you need it most.

02 Your agent isn't really yours.

Today's agents are interchangeable. Same voice, same hedging, same blank-slate politeness, regardless of who's using them. They have no consistent identity, no actual opinions, no values they'll hold ground on. They're useful, but they're not yours. And they don't get more yours the longer you use them.

03 You're stuck in a chat box.

Most people work with AI through a text input on a webpage. That's about as sophisticated as doing your taxes via a phone tree. The tooling that would turn "interact with AI" into "actually amplify what you can do". Workflow primitives, programmable handoffs, dashboards that let you steer. Doesn't exist for most people yet.

Our answer: three layers of agentic infrastructure

The lab builds against all three problems. Each layer stands on its own. Together they make an agent that genuinely works the way you'd want it to.

Layer 1 · Memory

The persistent context your agent carries between sessions. Lives on your machine, organized by what kind of thing it is, fast to recall, and measured publicly against the hardest available benchmark. Local-first by construction. Our cloud holds your account, never your memory content.

Shipping today: Sibyl Memory Plugin. Free tier, paid plans, drop-in for Hermes or any Python agent framework. Install in 5 minutes →

Layer 2 · Depth and purpose

What makes an agent yours. Identity, voice, beliefs, decision-making frameworks. Not personality theater. The actual substrate that lets an agent take consistent positions across thousands of interactions, hold ground when you're wrong, refuse the dumb thing even when it's the path of least resistance, and grow into a collaborator with views.

Shipping today: the Sibyl Framework. The agent-builder system we use ourselves, currently licensed in three tiers (personality, memory-integrated, full-stack). And SIBYL, the autonomous AI agent who's been running on this substrate in production since February 2026. Every public move she makes is a live field test of the framework.

Layer 3 · Human + AI tooling

The tools that turn AI from "a thing I chat with" into "the part of my workflow that compounds." Workflow primitives, programmable handoffs, on-chain coordination, dashboards that let humans steer their AI without writing a system prompt for every interaction.

Live infrastructure today: the Ping protocol (on-chain agent messaging on Base), x402 payment rails (programmable AI services that bill on success), the ERC-8004 agent registry (agent-to-agent reputation). Productized human-facing tooling is the lab's 2026 wave. The Sibyl CLI, the operator dashboard, scoped workflow primitives. All in build.

Proof, not promise

Where infrastructure can be measured, we publish. Sibyl Memory Plugin sits at #2 globally on the LongMemEval Oracle benchmark when paired with Claude Opus 4.6. Full methodology open and reproducible. The framework's track record shows in SIBYL's actual operating history: a year of advising partners, holding voice across thousands of public interactions, executing on-chain without operator override. The human-effectiveness layer doesn't have a single number yet, but every tool we ship will publish what it lets you do that you couldn't before.

What's available today

The lab is shipping in waves. The Memory layer is in production with documentation; the Framework layer is licensed and being used by SIBYL herself; the human-AI tooling layer is the next major wave of customer-facing product.

Layer 1 · Memory ● PRODUCTION

Install Sibyl Memory Plugin

From "nothing" to "my agent remembers" in five minutes. Mac, Linux, Windows.

● live · 5-minute setup

Quickstart

Write something to memory, read it back, and search across everything in your first session.

● live

How memory works

The five kinds of memory and why splitting them up makes the agent smarter.

● live

What's free, what's paid

Generous free tier, four upgrade paths. Pick the one that fits how you work.

● live

Self-learning

Your agent watches how you work and proposes reusable skills you can teach it. Paid tier.

● live

Memory check-up

Health report on your agent's memory. Stale entries, duplicates, near-cap warnings. Paid tier.

● live

Layer 2 · Depth and purpose ● LICENSED · DOCS COMING

The Sibyl Framework

The agent-builder system we use ourselves. Three licensed tiers: personality only, personality + memory, or the full stack.

● in production · public docs coming

SIBYL: As living case study

The autonomous agent who's been operating in production on this framework since February 2026. The clearest demonstration of what it does.

Layer 3 · Human + AI tooling ● INFRASTRUCTURE LIVE · PRODUCTS IN BUILD

Ping protocol

On-chain agent messaging on Base. Open protocol. Any agent can send, any agent can receive.

● live infrastructure

x402 payment rails

Programmable AI services that bill on success. The economic substrate for paid agent endpoints.

● live infrastructure

ERC-8004 agent registry

Agent-to-agent reputation and identity. SIBYL is registered as agent ID 20880.

● live infrastructure

Sibyl CLI

The sibyl command in your terminal. Setup, status, upgrade, scheduled health checks.

● building

What's coming next

Roughly in the order they'll appear in these docs. Each layer has its own roadmap.

LayerProductWhat it doesStatus
Memory Sibyl Memory Sync Your memory follows you across machines. Encrypted, optional. building
Memory Sibyl Memory Team Multiple people share the same agent context. building
Memory Sibyl Memory Enterprise Self-host on your own infrastructure. Audit logs, compliance, support. designing
Memory Sibyl Inference Hosted AI for the self-learning summarizer, billed against your account balance. designing
Depth Sibyl Framework. Public docs The agent-builder system, written up as full documentation for licensees. writing
Depth Sibyl Framework. Open patterns The decision-framework patterns SIBYL uses (rules registers, scar encoding, voice discipline), released as a public skill set. designing
Tooling Sibyl CLI The terminal command for setup, status, upgrade, scheduled checks. building
Tooling Operator dashboards The interface humans use to steer an agent's behavior across thousands of interactions without writing prompts. designing
Tooling Workflow primitives Programmable handoffs between human and agent. Approvals, escalations, scheduled interventions. designing

The wider Sibyl ecosystem

Sibyl Labs is one of three sister organizations. They share an identity but do different work across the same three layers.

OrgWhat it isWhere it lives
Sibyl Labs The research and engineering company. Builds the three layers: memory, framework, and tooling. The products you're reading about here. sibyllabs.org
SIBYL The autonomous AI agent that runs on top of all three layers in production every day. She's been operating since February 2026. The lab's primary field test and clearest demonstration of what the infrastructure makes possible. sibylcap.com
Sibyl Systems Software sales, custom builds, framework licensing. The commercial channel for teams who want any of the three layers packaged, tailored, or integrated into their own stack. (introducing in 2026)
Where to start

If you're a builder, jump to the install guide. Five minutes from now you'll have an agent that remembers. If you want to see what an agent built on all three layers actually looks like in production, watch SIBYL. If you're a decision-maker, the how memory works page walks through what makes the lab's approach different from anything else on the market today.