Infrastructure for autonomous agents

Build agents you can trust.

Memory that's proven, not guessed. Compliance that's enforced, not promised.

#2 on LongMemEval Install the top file-based memory plugin
Scroll to explore
The three things we build

Three products. One question underneath.

How does an autonomous agent earn the right to be trusted with real work? Everything we build is an answer to that.

  1. 01

    Sibyl Memory

    A memory that survives the session.

    Live

    Graph-structured, file-based memory an agent queries like a database, not a guess. #2 on LongMemEval Oracle. The highest-scoring file-based system on the leaderboard.

    pip install 'sibyl-memory-cli[mcp]' Read the docs
  2. 02

    Sibyl Sovereign

    Compliance you can prove.

    In development

    For agents doing work where a mistake is costly: moving money, touching customer data, making regulated decisions. You declare the rules the agent must follow, and a deterministic gate enforces them between the model and every action. A rule-breaking action never runs. The model can be wrong. The gate can’t.

  3. 03

    Living Graph Networks

    A living memory agents and people think in together.

    On the horizon

    Living Graph Networks is the next step past static storage: one shared, living memory that agents and humans plug into just as easily. Agent swarms can converse and ideate over the same data, people alongside them, so ideas compound and breakthroughs come faster. It keeps reshaping itself as it learns. Still file-based, still no vectors.

    Early research · no timeline yet

Same substrate underneath all three. Memory is live. Sovereign is being proven. Living Graph Networks is where the architecture goes next.

The record

Proven, not guessed.

Every number here is published and independently checkable. No self-reported ceilings.

LongMemEval Oracle
#2 on the leaderboard
95.6% native architecture
95.1% plugin · Sonnet 4.5
Independent 500-company benchmark
100% retrieval accuracy
1st of 4 engines tested

A 365-day simulated business year: 2,000 entities (500 companies, 1,500 people), 191k records, 350 questions posed identically to every engine. Sibyl retrieved 350/350 on an average of 2 rows per query.

Independently run at agent-bench.xyz
Built on

Infrastructure the enterprise already trusts, and our own product underneath it all.

Start here

Build agents you can trust.

Memory is live and public today. Install it in about two minutes, or browse everything we build.

Partnerships, bespoke deployments, and research collaborations: [email protected]