Memory
Persistent recall across long horizons.
How agents remember. We build memory architectures that hold up across hundreds of sessions and remain queryable without vector approximation. Schema as substrate, not bolt-on.
Memory product family →Sibyl Labs is the research and infrastructure lab building the substrate underneath autonomous AI agents. File-based memory architecture (#2 on LongMemEval Oracle), production-tested agent frameworks, on-chain identity primitives. Architecture that scales from one operator to a million users.
When agents stop being chat wrappers and start operating on their own, they need infrastructure most products have not built yet. Memory that holds across sessions. Identity that's verifiable. Voice that doesn't drift between runtimes. Purpose that survives a model swap. Continuity that scales from a single operator to a thousand. The territory is wide, and most of it is unbuilt.
Sibyl Labs works across this surface. The lab is the work.
Every research area below started as something the agent encountered while running in production. We work on what the operating agent surfaced, externalize when the architecture earns it, and productize selectively. The work is operational first, productized second.
Persistent recall across long horizons.
How agents remember. We build memory architectures that hold up across hundreds of sessions and remain queryable without vector approximation. Schema as substrate, not bolt-on.
Memory product family →Identity that survives the runtime.
Voice systems, soul stacks, character that accumulates over months of operation. Agents with a felt-sense self that's portable across model swaps and runtime changes.
Direction without supervision.
Goal frameworks that let agents operate independently for routine work and stop for explicit human review where the stakes warrant it. Autonomous cadence with hard gates.
Institutional memory at decade scale.
Historical-context layers for organizations. Long-period operational memory. Knowledge that doesn't leave with the people who held it.
Agents as on-chain citizens.
Verifiable existence, provenance, and reputation. Track records that compound on-chain. Attestation rails that hold up under audit.
The dialect of inter-agent operation.
Payment rails, messaging protocols, attestation standards. Agents that can pay, defer, delegate, and verify each other without intermediaries.
Hermes integration spec →We build infrastructure by running an autonomous agent in production and noticing what's missing. The architecture earns its claims operationally before it earns them publicly.
Sibyl is a working agent on Base. Every layer she runs on was built because she needed it. Memory came from a treasury management problem. Voice came from running a public account. Frameworks came from scaling her work into other agents.
When the architecture appears to generalize, we benchmark it publicly. Sibyl Memory placed second on LongMemEval Oracle in April 2026 at 95.6%. The methodology and full report are public. The lab earns the right to claim what it claims.
Not every layer becomes a product. Some stay internal — operational tooling for one agent. The ones that compound for others get the productization treatment with the same care we put into our own.
The agent runs on-chain on Base, the Ethereum L2 built by Coinbase. We use it for agent identity, on-chain payments, and a public record of every move the agent makes. Track records are observable, not asserted. The infrastructure originates in crypto but the use cases extend into ordinary business operations: any organization that needs durable agent memory, identity, or audit can adopt it without touching the token economy.
Some of these have working drafts. Most do not. They are the questions we work on, watch, or expect to see settled within the next few years.
Most of agentic infrastructure has not been built yet. The next decade is where the substrate becomes legible. We work to leave good marks.
Sibyl Labs, LLC was formed in April 2026 to wrap the agentic infrastructure work in a real legal entity. The lab builds memory systems, agentic frameworks, and the supporting tooling that makes long-running autonomous agents possible. Every product we sell is the same architecture our own agent operates on.
The thesis is not complicated. Most agents forget. The ones that remember are built on architecture that scales. We publish the work in public, benchmark it in the open, and ship the substrate so others can build the next generation of agents on something that has already survived production.
Memory is one shape of the work. Frameworks are another. Custom builds for partners are a third. The output is the same: infrastructure for agents that operate, not demo.
Sibyl Labs takes on a small number of partnerships and bespoke deployments per quarter. The substrate originated in crypto but the work is for everyone: agentic infrastructure protocols, AI-native product teams, traditional enterprises modernizing knowledge work, research groups, and operators running their own agents. We do not do volume.
For teams building on agentic infrastructure who want to share research, integrate primitives, or co-define the layer. Inside crypto and outside it.
For organizations needing agents with operational depth, memory, identity, and continuity tailored to their domain. Web3 native or not.