In 2026, every app remembers you. Your assistant imports your history from the last one, every tool
quietly assembles a profile, and the memory you generate becomes someone else's asset. The question
stopped being whether it remembers and became who the memory works for.
Sibyl is cross-agent memory where the answer is you. One knowledge graph holds your decisions,
your gotchas, and the conventions you actually follow. It is self-hosted on your own hardware and
shared across every coding agent you run: Claude Code, Codex, Cursor, and the agents you build
yourself. Tell one tool your context, and every tool keeps it. Nothing harvested into a vendor's
profile, nothing trapped behind an export you can't take.
Cross-agent memory for AI coding tools: one self-hostable knowledge graph shared by Claude Code, Codex, OpenCode, Cursor, and the agents you build.
스타
68
도구
—
가동 시간
—
The interaction surface is the shell. If your agent can run a command, it already speaks Sibyl. No
SDK to adopt, no MCP server you are locked into. The depth lives in the CLI; MCP is there for the
clients that prefer it.
Switch tools, swap models, walk to a new machine. The graph comes with you, because it was always
yours.
✦ What You Get
Capability
What It Means
🔮 Compounding Context
Every session adds to the graph instead of starting over. The longer you use it, the sharper it gets
🪄 The Memory Loop
context → act → remember → reflect runs through the CLI, skills, MCP, and hooks. Agents wake with context and leave it behind
🎯 Semantic Search
Find knowledge by meaning. "Authentication patterns" surfaces OAuth notes even when "OAuth" isn't in the text
🦋 Task Workflow
Plan with epics and tasks, then track execution across sessions and teammates in one place
🧪 Source-Grounded Synthesis
Draft verified documents from your own memory with citation, freshness, and gap checks
🌊 Source Ingestion
Crawl docs, import sources (mailboxes, archives), and ingest agent transcripts into scoped raw content and graph memory
💎 Scoped Multi-Tenancy
Namespace-isolated graphs, org-scoped content/auth records, and policy gates for personal, project, and team scopes
⚡ Quickstart
Connecting a machine to a Sibyl server takes one line: install the CLI, then run sibyl setup with
the server's URL.
sibyl setup signs you in through the browser, installs the Sibyl skill for
Claude Code, Codex, and other agents, and adds a Claude Code SessionStart hook. You can also hand
the job to your agent:
Set up Sibyl on this machine by following https://sibyl.example.com/agent
Copy the exact sentence from the web app's Connect card: /agent is served by the web app, so on a
deployment where the web app and API have separate origins the card points at /api/setup/agent.md
on the API instead.
No Server Yet? Run One Locally
# Shell installer: puts sibyl on your PATH, starts the local stack, opens setup UI
curl -fsSL https://raw.githubusercontent.com/hyperb1iss/sibyl/main/install.sh | sh
# Or Homebrew
brew install hyperb1iss/tap/sibyl && sibyl up
Finish the setup wizard in the browser, then run sibyl setup to connect this machine to the local
server. Use --daemon for the embedded daemon without the web UI. For containers,
sibyl docker init && sibyl docker up brings up API, web, and SurrealDB. Windows and the other
paths are in Installation.
First Five Minutes
# Capture a learning the moment you find it
sibyl remember "Stale auth token bug" \
"Redis TTL mismatch dropped the cached token early" --kind error_pattern
# Pull it back as working context for your next session
sibyl context "auth token bug" --intent debug
# Load a broader context across accessible projects
sibyl context "review stale auth token and Redis TTL history" --intent review --all
# Package wake-up context for the next coding session
sibyl session bundle
🪄 The Memory Loop
Sibyl is built around a durable loop that both humans and agents follow:
Context before you start. sibyl context "<goal>" returns a compact context pack: active
work, decisions, plans, constraints, and recent lessons, scoped to your project.
Act with that context in hand.
Remember durable knowledge as you learn it. sibyl remember stores decisions, plans, ideas,
claims, procedures, and gotchas so the next session does not rediscover them.
Reflect at clean breakpoints. sibyl reflect distills raw session notes into reviewable
memory candidates and can persist them into the graph.
sibyl context "ship the context graph" --intent build
sibyl remember "Use context packs" "Group memory before dispatching agents" --kind decision
sibyl reflect "We decided X. Next we build Y." --title "Planning checkpoint" --persist
Memory is graded, auditable, and scoped. Raw captures stay verbatim, reflection candidates pass an
automatic review before promotion, and a nightly dream-cycle keeps the graph consolidated. See
Capturing Knowledge.
The CLI
The CLI is the power-user interface: clean output, built for scripting and durable project
workflows.
# Memory loop
sibyl context "<goal>" # Compile working context
sibyl remember "Title" "Body" # Store durable memory
sibyl reflect "<notes>" --persist # Distill notes into candidates
sibyl capture "<quick note>" # Fast verbatim capture
sibyl correct <raw_memory_id> # Inspect or correct source memory
# Knowledge & graph
sibyl remember "Redis pooling" "Pool size must be >= concurrent requests" --kind rule
sibyl explore related ent_xyz # Find connected entities
sibyl show <id> # Full content by ID
# Task workflow
sibyl task list --status todo,doing
sibyl task start <task_id>
sibyl task complete <task_id> --learnings "Key insight: check TTL first"
# Synthesis, sources, projects, orgs
sibyl synthesis draft "Onboarding guide"
sibyl crawl add "https://docs.example.com" --name "Example Docs"
sibyl ingest claude-code ~/transcripts/ # Import agent transcripts into memory
sibyl docs list # Browse document collections
sibyl project link proj_xxx
Tasks flow backlog → todo → doing → review → done → archived (with a blocked side state). Every
list command supports --json for scripting, and the task, project, and entity lists also take
--csv. Full command reference: docs/cli/.
Graph: Interactive force-directed visualization of knowledge connections
Search: Semantic search with filters
Memory: The memory workspace, raw captures, source imports, and synthesis
Sources: Configure and inspect documentation crawling
Settings: Organizations, teams, API keys, security, LLM routing, backups
Built with: Next.js 16, React 19, React Query, Tailwind CSS 4, and the SilkCircuit design
system.
🛠️ Skills & Hooks
Skills are how an agent learns to use Sibyl, the bridge that makes "if your agent can run a
command, it speaks Sibyl" real. The loader teaches the memory loop, every CLI verb with the flags
that actually exist on your machine, context-pack usage, and the error patterns to avoid, so agents
stop guessing the interface from stale training data.
Skills are not Claude-only. The loader goes into every agent skill root Sibyl knows: Claude Code
(~/.claude/skills), Codex (~/.codex/skills), and the generic ~/.agents/skills convention. The
same workflow follows you across tools. sibyl setup installs it for you; the skill commands
manage it on their own.
sibyl skill install # Install the tiny /sibyl loader into every agent skill root
sibyl skill list # List the version-matched packs the CLI can serve
sibyl skill get core # Print the full workflow + command contract
moon run hooks:install # Optional Claude Code context hooks for repo dev
The installed skill is deliberately tiny: a loader that points the agent back at the CLI. The real
skill packs (contract, core, quick, workflows, examples, migration) are built into the
CLI and served on demand with sibyl skill get, each matched to the exact Sibyl version on the
machine. Upgrade the CLI and the guidance upgrades with it. No stale skill copies drift out of sync,
and a subagent on any host gets the same source of truth from one command.
Hooks are separate and, for now, specific to Claude Code: a single SessionStart hook prints a
compact wake-up bundle with active tasks and relevant memory, then the agent owns invoking the
sibyl skill and calling sibyl context for working context. sibyl setup registers it when
Claude Code is installed (--no-hooks skips it). See Skills & Hooks.
MCP Integration
Most agents need no MCP config: they run the sibyl CLI, and sibyl setup already taught them how.
For a client that only speaks MCP, such as Cursor or Claude Desktop, create a key with
sibyl auth api-key create --name cursor --scopes mcp and point the client at your server's /mcp
endpoint:
Sibyl benchmarks LongMemEval-S retrieval on the live /api/search path, in an ephemeral CI stack,
with no LLM extraction and no LLM reranking. No retrieval number is published right now: the pre-1.0
headline run was withdrawn, and a new number lands here only once a live run on a current release
passes the benchmark gate.
hit@5 and strict recall@5 measure different things, and many published "LongMemEval" numbers are
end-to-end QA accuracy with an LLM judge, a different metric than retrieval recall. Methodology and
claim boundary: LongMemEval-S ·
AI Memory Landscape.
Authentication supports JWT sessions for the web UI, scoped API keys for programmatic access
(sibyl auth api-key create --scopes mcp,api:read), GitHub OAuth, enterprise OIDC/SSO with exact
provider-to-organization binding, and self-service password reset over SMTP. MCP endpoints enforce
Bearer auth when a JWT secret is set. Detailed guides:
Docker Compose · Kubernetes ·
Environment · Auth reference.
Knowledge Model
Sibyl models a wide range of entity types so memory stays structured, not just a pile of notes:
The full registry and how types relate live in Entity Types.
FAQ
Who is Sibyl for?
Solo developers who want durable memory for projects and debugging. Teams who want shared
knowledge that compounds. Anyone building with AI who wants the memory they generate to stay
theirs, across every tool.
Do I need AI agents to use Sibyl?
No. The knowledge graph and task system are the core product: documentation, task tracking, captured
learnings, and semantic search over what your team already knows. AI agents make the memory loop
automatic, but they are not required.
How does it compare to Mem0 / LangMem / similar?
Sibyl is self-hosted and open source, so you own your data. Where most memory products stop at
recall, Sibyl ships a full task workflow system and a web UI for the humans on your team, and it
keeps knowledge, tasks, and docs connected in one graph instead of scattering them across tools. On
retrieval quality, see the Benchmarks above.
What LLM APIs do I need?
A language model (required): Anthropic, OpenAI, Gemini, or Claude through Amazon Bedrock, for
surfaces such as crawler extraction, synthesis, and reflection.
An embedding provider (required): OpenAI, Gemini, or Cohere Embed v4 through Amazon Bedrock,
for embeddings and semantic search.
Bedrock signs requests with AWS credentials, so it needs no provider API key; see
Amazon Bedrock. The server holds these settings, so
on a shared server the operator configures them once and nobody connecting with sibyl setup enters
a key. Providers and models are configurable globally or per surface from the web admin settings. A
typical solo developer uses around $5/month in API costs.
Is it production-ready?
Yes. Sibyl is 1.0. SurrealDB is the runtime for graph, content, and auth. We use Sibyl to
build Sibyl. Every feature, task, and learning you see here was tracked inside the system itself.
🎯 What's Next
Past 1.0, the work is making memory more automatic and the sources deeper:
Memory Workspace OS: automatic memory decisions become visible, explainable, correctable, and
undoable from one product surface.
Correction propagation: edits, redactions, and rollbacks flow forward into future recall and
synthesis instead of stopping at the record you touched.
Broader ingestion: source import reaches past mailbox-style archives into more corpora, scoped
and provenance-tracked.
Already shipped on the way here: a Surreal-only default runtime, plus source, document, and agent
transcript ingestion via sibyl ingest and sibyl docs. See
docs/architecture/SIBYL_NORTHSTAR.md for the full
direction.
💜 Philosophy
Recall before you act. The graph knows things, so pull context before you code. Work in task
context for traceability, progress, and knowledge linking. Remember what you learn: if it took
time to figure out, save it so the next session doesn't pay for it twice.
Bad: "Fixed the bug." Good: "JWT refresh fails when Redis TTL expires. Root cause: token
service does not handle WRONGTYPE. Fix: try/except with regeneration fallback."
The graph should be smarter after every session.
Contributing
We welcome contributions. See CONTRIBUTING.md for guidelines.
# One-line setup (installs proto, moon, toolchain, dependencies)
./setup-dev.sh # macOS / Linux
pwsh -File .\setup-dev.ps1 # Windows (PowerShell 7+)
# Configure your shell, then launch the local-dev stack
export SIBYL_OPENAI_API_KEY=sk-... # embeddings; SIBYL_JWT_SECRET auto-generates in dev
moon run install-dev # editable CLI installs
moon run dev # SurrealDB + API + web
# Verify, then run the quality gates before a PR
curl http://localhost:3334/api/health
moon run :check # lint + typecheck + test
moon run dev is the single-machine flow: with SIBYL_SURREAL_URL unset it starts local SurrealDB,
points the API at ws://127.0.0.1:8000/rpc, and stores data in .moon/cache/surreal-dev. Set
SIBYL_SURREAL_URL to use a hosted endpoint, including Surreal Cloud.
이 페이지에서 Cursor, Claude Desktop, Claude Code 또는 VS Code용 stdio 설정을 복사합니다. PlugMCP 프록시가 아니라 게시자의 설치 방식을 사용합니다.
Sibyl에 API 키가 필요한가요?
이 레코드에는 필수 환경 변수가 등록되어 있지 않습니다.
Sibyl은(는) 어떤 전송 방식을 사용하나요?
Sibyl은(는) stdio를 통해 로컬에서 실행됩니다.
Sibyl은(는) 마지막으로 언제 확인되었나요?
마지막 확인: 2026-11-13. Cross-agent memory for AI coding tools: one self-hostable knowledge graph shared by Claude Code, Codex, OpenCode, Cursor, and the agents you build.