# Sibyl-Memory

A durable, file-based long-term memory system for AI agents distributed as five packages, an SDK, a CLI, an MCP server, a Hermes adapter and a LangGraph store, with no vector database and no embeddings.

Sibyl-Memory is a local-first agentic memory SDK backed by SQLite, using a five-tier hierarchical schema and full-text search rather than a vector database or an embedding model; the memory itself lives in a single file on the agent's machine. It is multi-tenant by design and includes self-learning skill detection and a local memory linter.

Four further packages build on the SDK: a CLI for activation and tier management, an adapter for the Hermes Agent runtime, an MCP server for any MCP-compatible client such as Claude Code, Codex, Cursor or Continue, and a LangGraph BaseStore adapter. The only outbound network call in normal use is a tier-verification check that sends account metadata, never memory content, when an activated account writes past its tier cap; unactivated, free use makes no network calls at all.

According to the project, the architecture was benchmarked on the LongMemEval Oracle evaluation and placed second overall, the only file-based system in the top tier, running on a small four-vCPU, 16 GB machine. It runs on Linux, macOS and Windows through WSL2, and installs from PyPI.

Website: https://sibyllabs.org/plugin