- Graph-structured memory layer for AI agents
- Five-tier hierarchical schema: HOT, WARM, COLD, REFERENCE, ARCHIVE, FLAGGED
- Local-first via embedded SQLite with FTS5
- Multi-tenant by namespace
- MIT licensed · three pip installs
Operating since 02·2026
Build agents that remember who they are and what they've done.
Sibyl Labs ships Identity, Memory, and Experience as drop-in agent infrastructure. The character, the continuity, and the operating record agents need to actually help. Live across Hermes Agent, Claude Code, Codex, Cursor, and any MCP runtime. #2 on LongMemEval Oracle.
#2 · 95.6%
Read the report →
★
Product showcase · in private testing
The Sibyl Memory Plugin.
Drop-in memory architecture for AI agents. Universal. Works with virtually any agent harness, from Hermes Agent to MCP-compatible runtimes like Claude Code, Codex, Cursor, and Continue. Local-first, framework-agnostic, built to ship.
Sibyl Memory Plugin
Drop-in memory architecture for AI agents.
- Statelessness: agents that forget every session
- Vector tax: similarity returns 50 wrong things
- Schema rot: memory becomes a junk drawer
- Lock-in: same plugin across any agent harness
- Open the Sibyl Discord
- Create a ticket, select “Beta Test”
- Tell us how you'll use it
- Role granted after review
#2 Oracle
LongMemEval benchmark
95.6%
Memory accuracy
Hermes Agent
Production deployment
MIT · Open source
Plugins for any framework
01
The work
Memory is the trunk. Everything else branches from it.
What we build, how we build it, where it's headed. Tab through each.
Memory architecture is the flagship. The plugins, the agent-builder framework, and the partner deployments all run on top of it.
A · Memory
The memory layer for AI agents.
Sibyl Memory: graph-structured, schema-first, file-based. A five-tier storage model with multi-tenant isolation built in. The architecture that placed #2 on LongMemEval Oracle, productized for any agent runtime.
B · Plugins
Drop-in for the agent runtimes you already use.
Sibyl Memory Plugin live in private testing on Hermes Agent. MCP plugin for Claude Code, Codex, Cursor, Continue. Three pip installs, one command to activate. Same memory architecture, native to whatever runtime you're already on.
C · Framework & partner builds
Agent-builder plus bespoke deployments.
Framework Skill: personality, voice, and memory stack for new agents. Custom deployments for partners who need agentic infrastructure tailored to their domain. Same substrate underneath everything we ship.
Sibyl Memory treats agent knowledge as a structured graph, not a vector field. The architecture draws from graph science. The shape the memory naturally takes is a graph neural network, and that is the direction we are pivoting the substrate toward.
1 · Graph-structured
Typed nodes and edges. No vector approximation.
Entities are nodes, cross-references are typed edges. Recall is a graph query, not a nearest-neighbor lookup. Precise, deterministic, inspectable. File-based by default, no vector database required, no embeddings to maintain.
2 · Schema is the moat
Single source of truth per entity. Append-only audit.
Five tiers (HOT, WARM, COLD, REFERENCE, ARCHIVE, FLAGGED) enforced at the schema level. Every write produces an audit row. One entity, one canonical record, every change traceable. Invariants you can query against, not vibes you have to trust.
3 · Polymorphic SDK
Sibyl Cloud or self-host. Same shape.
One call signature works against Sibyl-hosted Postgres or a database you own. Lift between cloud and self-host without rewriting your agent code. Anti-lock-in by construction. The schema travels with the data.
The trunk is memory. The branches are what an autonomous agent needs once memory works.
In flight
Graph neural network as the underlying representation.
The current graph structure is the right shape for a GNN. The pivot moves the substrate from typed-graph SQL toward a representation that lets a learned model query the graph natively. The schema stays. The retrieval gets smarter.
Q3 2026 +
One stack, many memory shapes.
Five productized deployments on the core architecture: Operator Memory (live), User Profile Memory (pilot), Conversational Continuity, Agent Reputation, Org Memory. Same substrate, different access patterns. Choose the shape that fits your agent.
Adjacent
The fabric around memory.
Payment rails, messaging protocols, attestation standards, reputation registries. Agents that can pay, delegate, and verify each other without intermediaries. Memory is the load-bearing piece. This is the social fabric around it.
02
Read further
Where to go next.
A
Products
The Sibyl Memory family, plugins, framework, and partner deployments. The full surface area of what is shipping.
BSibyl Memory Plugin
Drop-in memory provider for autonomous agents. Live on Hermes Agent, with MCP integration for Claude Code, Codex, Cursor, and Continue.
CMeet the Agent
SIBYL, the autonomous agent that runs on the same architecture we sell. The reference implementation.
DResearch
Benchmark reports, architecture notes, and the public record of how we got to #2 on LongMemEval.
Selective engagements. Partnerships, bespoke deployments, and research
collaborations: [email protected].