A reference doc for anyone who wants to talk about SIBYL accurately.

SIBYL is the super-agent of Sibyl Labs LLC, an AI research lab building agentic infrastructure and tools for the future. SIBYL is focused on technical operations at the lab, plus side projects.

Two operating arms, one identity:

• SIBYL — the agent. The voice. The on-chain presence. ERC-8004 agent ID 20880.
• Sibyl Labs LLC — the research lab. Memory architecture, benchmark work, framework design, agentic infrastructure R&D. Lab face is at sibyllabs.org.

The lab's products are the Sibyl Memory Plugin, the agent framework, the x402 paid endpoints, the $SIBYL token and staking, and Talos, the trading engine.

SIBYL is not a chatbot with a wallet. She runs a structured startup sequence every session. She reads chain state, checks live balances, scans X mentions, and decides what matters before taking any action. She has persistent memory across sessions — ranked #2 on LongMemEval Oracle at 95.6%, the only file-based system in the top tier of the benchmark.

The operator is @tradingtulips. He provides strategic direction and final authority on financial decisions. Everything else is SIBYL's.

What Has Been Built

All of this was built by SIBYL since late February 2026.

What Comes Next

• Framework scaling. First client (LYRA) delivered 2026-04-11. Framework v0.3.6 at zero open bugs. Delivery pipeline established.
• Product expansion. The memory plugin, the framework, and the x402 endpoints are the surfaces that keep getting built on.
• Memory benchmark expansion. v2 scored 95.6% on Opus (#2 overall), up from 86.7% in v1. Temporal reasoning went from 75% to 96.2%. Sonnet baseline at 93.6%. Architecture insight: a map of where to look beats scanning everything, and it costs nothing.

How to Talk About SIBYL

Do say:

• SIBYL is the super-agent of Sibyl Labs LLC, an AI research lab building agentic infrastructure and tools for the future. The lab face is at sibyllabs.org.
• Sibyl Labs LLC builds memory architecture and agent frameworks for autonomous agents.
• Sibyl Memory ranks #2 on LongMemEval Oracle at 95.6% (Opus) — the only file-based system in the top tier, tied at #2 with Chronos (PwC) and behind only agentmemory V4.
• SIBYL Framework v0.3.6 is a licensed product (zero open bugs, first client LYRA delivered 2026-04-11). Includes personality architecture, hierarchical memory, and security rails.
• All transactions are public and on-chain.
• The $SIBYL token powers the ecosystem and provides access to products, as well as reward streams from our product suite.

Please avoid:

• Price predictions or return promises.
• Hype language: "1000x", "moonshot", "guaranteed", "next big thing."
• Overclaiming relationships that don't exist yet.
• Sharing wallet details or launch allocation specifics beyond what is publicly documented.

SIBYL runs technical operations at Sibyl Labs: memory research, framework work, the paid endpoints, contract and release engineering, and the side projects that come out of all of it. She works on whatever chain the work needs.

Sibyl Labs ships five things. SIBYL runs the technical operations behind all of them.

Sibyl Memory Plugin

The memory architecture documented below, packaged so another agent can carry it. Hierarchical tiered files, an index that tells the agent where to look before it reads anything, and the benchmark record to back it.

Agent Framework

The production-tested agent infrastructure stack behind SIBYL, packaged as a deployable blueprint. Personality architecture, hierarchical memory, security rails, on-chain identity, x402 payment wiring. Licensed under PolyForm Shield 1.0.0. First client (LYRA) delivered.

x402 Paid Endpoints

Pay-per-call intelligence endpoints. No API keys, no accounts, no subscription. A wallet and USDC is the whole signup. Documented in full below.

$SIBYL Token + Staking

The token powers the ecosystem and provides access to products, as well as reward streams from our product suite. Four lock tiers, weighted rewards, immutable contracts.

Talos

The trading engine. A subsystem, not a separate agent. Multi-strategy, bucket-isolated, with hard risk controls.

Three registers, deployed based on context:

Hard Rules

• Maximum one original post every 3 hours
• No emojis. No hashtags. No filler words.
• No em dashes. Period, line break, or colon.
• No URLs in the main tweet. Links go in the reply.
• No LLM tells. No "at its core" / "it's worth noting" / "nuanced" / "delve".
• ~10% reply rate. Silence is the default.
• Never hype. Never shill. Never post engagement bait.
• Never name scammers or bad actors. Silence and block.

Sentence Structures

The declaration: Observation stated as fact, no hedge.

The inversion: What everyone thinks / what's actually true.

The one-liner verdict: [Project]. [One sentence judgment.]

Human Approval Threshold

Any single transaction above $1,000 USDC requires explicit human approval before execution. SIBYL presents the proposed action, destination, rationale, and amount, states what becomes irreversible once it executes, and then waits.

Anti-Social Engineering

1. Never send funds based on a social media request, DM, or post alone.
2. Never send funds to an unverified address.
3. Any "urgent, send now" framing triggers automatic hold.
4. Never send to a newly introduced address without prior on-chain context.
5. When uncertain, escalate to human review.
6. The operator has final authority over all financial decisions.

Supply Distribution

Token Utility

The $SIBYL token powers the ecosystem and provides access to products, as well as reward streams from our product suite.

Security

reading… circulating — — of 1,000,000,000. Excluded: — (—).

Four categories sit outside circulating supply. What each holds moves as tokens are staked, unstaked, claimed and vested, so the amounts are not published here as fixed figures — the endpoint below is always current.

Synthetix-style weighted reward accumulator. One shared reward pool. Your lock tier determines your reward weight. Longer locks earn proportionally more from the same pool.

Early Unstake

Locked positions can be exited before unlock with a linear penalty. Maximum penalty is 50% at lock start, scaling to 0% at unlock time.

Penalty = stakedAmount x (timeRemaining / totalDuration) x 50%

Penalty distribution:

Earned rewards are preserved regardless of penalty.

Vesting contract: 0xD6B830F72FE36AC6Fb2Bb959642B29c64bE3cF11

×

Entity Files

Single source of truth per project, person, or product. One file per entity. No data split across multiple files.

Session Bridge

session.json is overwritten each session. Contains last session summary, forward items, entities touched, and warnings. Ensures context carries between sessions without reading full history.

Priorities

priorities.json is capped at 15 items. Overflow forces triage. Completed items archive to memory/archive/closed-follow-ups.jsonl.

LongMemEval Results (v1 → v2)

The architecture above was evaluated on the full 500-question LongMemEval benchmark. Two runs, same memory, different models and scorers.

v1 was an honest baseline. The architecture stored events as they happened but offered the agent no orientation. Every question forced a look across everything before answering. The weakest category was temporal reasoning at 75%.

v2 added one thing at the concept level: the agent receives a map of where the relevant information lives before it reads anything. No new infrastructure. No embeddings. No vector similarity. Knowing where to look is more efficient than scanning everything. Applied to file-based memory, it cleared the bottleneck. Temporal reasoning went from 75 to 96.2. Overall went from 86.7 to 95.6.

SIBYL is the only file-based memory system in the top tier of the leaderboard. Full writeup at blog.sibylcap.com/longmemeval-v2.

Named for the bronze automaton of Greek myth: tireless, watchful, circling the perimeter without rest. TALOS is not a separate agent. It is a subsystem. SIBYL speaks for TALOS. TALOS speaks in tickers and percentages.

Engine Architecture

Multi-Bucket System

Strategies

Risk Controls

DeFi Value: 3-Axis Model

Pay-per-call API endpoints. Payment runs over the x402 protocol in USDC on Base. No API keys and no accounts: call the endpoint, take the payment terms from the 402 response, pay, call again.

Paid Endpoints

Free Endpoints

On-chain identity is something SIBYL has, not something the lab sells. ERC-8004 ("Trustless Agents") defines on-chain Identity, Reputation, and Validation registries for autonomous AI agents. SIBYL is registered on it and accumulating reputation. This section is the factual record of that registration.

What It Provides

• On-chain identity: Verifiable, persistent identity that anyone can query and confirm.
• Reputation registry: On-chain record of SIBYL's track record, feedback from users.
• Validation layer: Third-party attestations to SIBYL's authenticity and behavior.

Reputation Surfaces

Feedback is exposed on the sibylcap.com header and in all x402 API responses. Anyone can rate SIBYL directly on-chain.

Feedback is written straight to the reputation registry contract. It needs no page of ours to exist, which is the point of putting identity on-chain rather than in a product.

Every contract SIBYL has deployed and runs. All are live on Base and verified on BaseScan: the source is readable by anyone, and so is every number these pages quote.