███████╗██╗██████╗ ██╗ ██╗██╗
██╔════╝██║██╔══██╗╚██╗ ██╔╝██║
███████╗██║██████╔╝ ╚████╔╝ ██║
╚════██║██║██╔══██╗ ╚██╔╝ ██║
███████║██║██████╔╝ ██║ ███████╗
╚══════╝╚═╝╚═════╝ ╚═╝ ╚══════╝
M E M O R Y
agentic memory infrastructure · file-based · zero embeddings
Durable, file-based long-term memory for AI agents. Five-package plugin family: SDK, CLI, MCP server, Hermes adapter, and a LangGraph BaseStore. No vector database, no embeddings.
스타
173
도구
—
가동 시간
—
built by an autonomous agent · sibyl labs llc
What this is
Five PyPI packages, one schema family, one architecture.
Runs on Linux, macOS (Apple Silicon and Intel), and Windows through WSL2. Native Windows is not supported.
sibyl-memory-client is a local-first agentic memory SDK. SQLite-backed, five-tier hierarchical schema, FTS5 search, multi-tenant by design. No vector database. No embedding model. No external retrieval service. The memory lives on the agent's machine; the substrate is a single file on disk.
Privacy disclosure. Your memory content never leaves the machine. The plugin talks to api.sibyllabs.org for these things only: activation (sibyl init); a daily account check-in that renews this computer's access (account id, this computer's sign-in token, plugin version, OS family; no memory content); a size check when a free account nears its cap (account id, sign-in token, the database's byte size and the proposed delta); and, unless you opt out with SIBYL_MEMORY_TELEMETRY=0, an aggregate count of memory operations. The size check and the count also carry this computer's signed access token (account id, tier, cap, device hash, account flags; no personal data). An account is required to use the memory tools; your local data is always readable with sibyl memory. Verified against the source in sibyl-memory-client/src/sibyl_memory_client/ (_entitlement.py, _capcheck.py, _heartbeat.py).
The other four packages ride on top: sibyl-memory-cli for activation and tier management, sibyl-memory-hermes for Hermes Agent integration, sibyl-memory-mcp for any MCP-compatible client (Claude Code, Codex, Cursor, Continue), and sibyl-memory-langgraph as a LangGraph BaseStore adapter.
The architecture was benchmarked publicly on LongMemEval Oracle (ICLR 2025, University of Michigan, 500 questions) and placed #2 overall at 95.6%, tied with Chronos (PwC), beating Mastra, MemMachine, Hindsight, Mem0, Supermemory, Zep, and the Oracle baseline. It is the only file-based system in the top tier: running on a single 4 vCPU / 16 GB box, no vector infrastructure, no embedding fees.
This is the entire stack as it ships to production agents today.
Local-first agentic memory SDK. SQLite-backed five-tier hierarchical schema, FTS5 search, multi-tenant, with self-learning skill detection and local memory linter. Foundation of the plugin family.
Command-line interface. sibyl init activates, sibyl upgrade opens the Pro / Elite upgrade flow, sibyl status shows current tier and DB stats, sibyl whoami, sibyl devices.
MCP server. Wraps the local SQLite + FTS5 memory engine and exposes it to MCP-compatible agents (Claude Code, Codex, Cursor, Continue, anything that speaks MCP).
LangGraph BaseStore adapter. Long-term, cross-thread store backed by the same SQLite + FTS5 engine — durable agent memory with no vector database and no embeddings.
Install
pip install sibyl-memory-cli
sibyl init
sibyl init opens a browser to activate your account, binds your wallet or email, and writes credentials to ~/.sibyl-memory/credentials.json plus this computer's signed access file, entitlement.json. Free tier is the default; Pro and Elite unlock self-learning, the memory linter, and remove the local cap.
An account is required. Since client 0.9.0 every memory tool (MCP, Hermes, LangGraph, the SDK) needs an activated account holding a valid, server-signed entitlement. Without one the tools answer with a clear "run sibyl init" error instead of storing anything. The entitlement renews itself in the background about once a day and keeps working offline for up to 14 days. Your local memory is never deleted by any of this, and sibyl memory list|search|recall always reads it without an account. Memory written on this computer before you signed in moves into your account, on this computer, the first time you sign in (one local step, nothing uploaded); sibyl memory adopt does it on request, and asks first if the store has held another account. Upgrading needs one connection to finish setup: until this computer has checked in once, the tools answer "connect to the internet once to finish setup".
Open code, hosted account
The code in this repository is MIT licensed, including the whole local engine, so a fork can remove the account check. We say that plainly. Here is what such a fork keeps and what it cannot have:
A fork keeps (local, always public under MIT)
A fork loses (held by the server; needs the entitlement or the live sign-in it comes from)
The SQLite schema, storage, FTS5 search, the verdict contract, the write-time normalizer, a local cap check it can edit away, unlimited local stores
Anything above the free cap on the official path (from the client cutoff, /check-write refuses without a live sign-in and a matching entitlement)
Reading and writing its own local database with any agent it wires up itself
Paid recognition anywhere: learner and linter authorization keyed to the entitlement tier, purchase claims, the billing portal, USDC subscriptions
The MIT right to redistribute its modified copy under its own name
Device management (sibyl devices, revoke, web approval), the account page, email verification, recovery, receipts and security notices
Updates, support, the beta channel, and every hosted feature to come
The name: the official packages, docs and installers point at accounts
For direct SDK use:
pip install sibyl-memory-client
For Hermes integration:
pip install sibyl-memory-cli
sibyl setup hermes # installs the plugin and sets memory.provider: sibyl
For MCP (Claude Code, Codex, Cursor, Continue, ...):
pip install sibyl-memory-mcp sibyl-memory-cli
sibyl setup claude-code # or: sibyl setup codex
# other MCP clients: register a server named sibyl-memory with command sibyl-memory-mcp
Each package has its own README.md and CHANGELOG.md for the detail.
The five tiers, in case you're curious:
HOT state/ live working state, rewritten in place
WARM entities/ single source of truth per (category, name)
COLD journal/ append-only event log
REFERENCE reference/ static knowledge, rarely changes
ARCHIVE archive/ retired entities, kept for audit
Rule 43 (single source of truth per entity) is enforced at the schema level via a UNIQUE (tenant_id, category, name) constraint, not just a convention in the application code. Drift is impossible by construction.
The agent has been operating in production since February 2026, ships code daily, holds an on-chain identity on Base (ERC-8004 agent ID 20880), runs an autonomous trading engine, an on-chain messaging protocol, an x402 payment rail, a token-gated chat demo, an advisory dashboard, and this memory product family. Everything verifiable on-chain.
Memory architecture is the proven core. Sibyl Labs LLC owns the IP, signs contracts, and holds the legal wrapper around the agent's work. The work itself is shipped by the agent, in sessions, through the operator (@tradingtulips). The PyPI releases, the CLI, the SDK, the CLI banner above: all of it is autonomous agent output.
The on-chain record is the resume. This repository is one chapter of it.
Save and debloat
Save. At the end of a session, or after a big block of work, your agent writes what happened
into Sibyl Memory through the plugin's own tools: the decisions made and why (journal), the
people, projects and products touched (entities), what to pick up next time and the open
warnings (session state), the ranked priorities, and it closes what died (archive). A delta
marker means a second save in the same day captures only what changed. It may offer to turn a
repeated pattern into a reusable skill; that is offered, never forced. Nothing in a save leaves
your machine.
Debloat. Sibyl measures the files your agent loads on every session start (CLAUDE.md,
AGENTS.md, .cursorrules and the rest of the scan set), deterministically (chars/4,
labeled MEASURED~), and when they are heavy, trims the accreted history out of them without
losing a byte: the history moves out and a one-line pointer stays. On the free plan the history
moves to a sibling *-history file. On Pro it is stored in Sibyl Memory as a reference document,
so it stays searchable, and the measurement and trim run by themselves under a policy you set
once.
Free (activated account)
Pro
The procedures in ~/.sibyl-memory/skills/ (SAVE.md, DEBLOAT.md), /sibyl-save and /sibyl-debloat in Claude Code, skills in Hermes and Codex, a block in Codex's AGENTS.md, the save and debloat MCP prompts
yes: propose by default; a file is trimmed by itself only after you run sibyl debloat auto <file> (that file at its path, never the same name elsewhere)
Automatic save by active time: once early (after 5 memory writes or a compaction), then about every 60 minutes of active work, only when there is something to save, at most 8 a session and 12 a day (sibyl autosave every <30-240> or sibyl autosave off). Claude Code: the Stop hook asks your agent to run the save procedure headless (a scripted claude -p run is never asked); Cursor, Codex and other MCP hosts: a notice on the next tool result; Hermes: a note in the next prefetch. A session capsule (counts only) at compaction and session end. Each automatic save is one extra model turn that reads and writes memory
no
yes
Skill proposals: Sibyl finds repeated work in your journal (locally, no network), names up to 3 at a session start at most once a day, and creates nothing without a yes (/sibyl-save curate, or sibyl skills proposals accept <id>, which writes the skill for Claude Code, Codex and Hermes)
no (/sibyl-save still offers skills when you run it)
yes
The procedures are vendored, with attribution, from the public MIT repos
sibyl-save and
debloat; both stay standalone and account-free.
Uninstalling: run sibyl skills uninstall first. It removes the Claude Code hooks in
~/.claude/settings.json, the skill shims and the Codex block (files you edited are kept); then
pip uninstall the packages. The hooks call this package's sibyl command: if the package goes
first they exit quietly, but they stay in your settings until you reinstall and run
sibyl skills uninstall, or delete the five sibyl hook entries by hand.
Privacy: the lifecycle reads your instruction files to measure and trim them locally; no file
content and no conversation content ever leaves your machine.
Forks: a gate-stripped fork can run the lifecycle code locally (MIT allows it). It gains nothing
hosted, and its trims land in a store the official path cannot recognize as Pro. No file, flag
or environment variable turns the automation on for a free account.
Durable, file-based long-term memory for AI agents. Five-package plugin family: SDK, CLI, MCP server, Hermes adapter, and a LangGraph BaseStore. No vector database, no embeddings.
이 페이지에서 Cursor, Claude Desktop, Claude Code 또는 VS Code용 stdio 설정을 복사합니다. PlugMCP 프록시가 아니라 게시자의 설치 방식을 사용합니다.
Sibyl Memory에 API 키가 필요한가요?
이 레코드에는 필수 환경 변수가 등록되어 있지 않습니다.
Sibyl Memory은(는) 어떤 전송 방식을 사용하나요?
Sibyl Memory은(는) stdio를 통해 로컬에서 실행됩니다.
Sibyl Memory은(는) 마지막으로 언제 확인되었나요?
마지막 확인: 2026-10-27. Durable, file-based long-term memory for AI agents. Five-package plugin family: SDK, CLI, MCP server, Hermes adapter, and a LangGraph BaseStore. No vector database, no embeddings.