Sibyl Memory

Local-first agentic memory. Your agent keeps durable memory across sessions in a single SQLite file on its own machine. No vector database, no embedding model, no external retrieval service.

Five PyPI packages, one schema family, one architecture. The same memory engine that placed #2 overall on LongMemEval Oracle at 95.6%, productized so any agent can run it.

What it is

sibyl-memory-client is the foundation: a small Python SDK that gives any AI agent durable memory, organized by what kind of thing it is rather than dropped into one fuzzy similarity bucket. The memory lives on the agent's machine. The substrate is a file on disk.

  • Local-first. Memory lives in SQLite under ~/.sibyl-memory/. No cloud round-trip for any read or write.
  • Organized by kind. Five tiers, each recalled the way it should be: working state, entities, events, reference, archive.
  • Zero embeddings. FTS5 full-text search across every tier. No vector index, no embedding fees, no retrieval model to host.
  • Multi-tenant. One machine can hold separate, isolated memory for separate identities.
  • Benchmarked in public. #2 on LongMemEval, the only file-based system in the top tier.
Privacy

Your memory content never leaves the machine. An account is required, and a signed-in computer checks its plan with your account (about hourly on Free, daily on Pro), makes a size check near the free cap, and sends an aggregate count of memory operations tied to your account. All of it is account metadata, never the contents of your memory. See Tiers & access for the exact behavior.

The packages

sibyl-memory-client

The local-first SDK. SQLite + FTS5, five-tier schema, multi-tenant. Everything else rides on top.

● on PyPI · SDK

sibyl-memory-cli

Command line. sibyl init signs you in, sibyl setup wires your agent and installs save and debloat, sibyl status shows your plan and access.

● on PyPI · CLI

sibyl-memory-hermes

Bundled memory provider for Hermes Agent and any Python orchestration wanting direct SDK access.

● on PyPI · provider

sibyl-memory-mcp

MCP server. Exposes the local memory engine to Claude Code, Codex, Cursor, Grok Bot, Continue, anything that speaks MCP.

● on PyPI · MCP

sibyl-memory-langgraph

A LangGraph BaseStore backed by Sibyl Memory: long-term, cross-thread memory for your graphs.

● on PyPI · LangGraph

Start here

Install the CLI and you have the SDK, the Hermes provider and the MCP server with it. Two commands, and the second one signs you in (an account is required; Free needs no card):

$ pip install 'sibyl-memory-cli[mcp]'
$ sibyl init

Sibyl Memory runs on Linux, macOS (Apple Silicon and Intel), and Windows through WSL2. Native Windows is not supported.

Then read Install for wiring it into your agent, How it works for the architecture, or CLI reference for the full command set.

Buying Pro for a group? Teams puts three or more Pro seats on one bill; each member keeps their own account and their own memory.

Why it holds up

The architecture was benchmarked on LongMemEval Oracle (ICLR 2025, University of Michigan, 500 questions) and placed #2 overall at 95.6% with Claude Opus 4.6, tied with Chronos (PwC), ahead of Mastra, MemMachine, Hindsight, Mem0, Supermemory, Zep, and the Oracle baseline. The productized plugin scored 95.1% on Sonnet 4.5, matching the architectural ceiling within half a point. Full numbers and methodology on the Benchmarks page.