Persistent memory for the AI you build with
_{✦ Yours alone · In your projects · In the systems you share ✦}

Why Sibyl? • Quickstart • Memory Loop • CLI • MCP • FAQ

Sibyl is persistent memory for the coding agents in your terminal, the projects you ship with them, and the teams you build alongside. One self-hostable knowledge graph, scoped by context. Yours to run, yours to keep.

Most AI sessions start cold. Coding agents forget yesterday's decisions. Notes drift between worktrees, tasks scatter across tools, and the context you earned the hard way evaporates the moment a session ends.

A durable knowledge graph holds what matters: personal context, project work, shared spaces, source documents, decisions, agent state. Each lives in its own scope and gets tied to the others when context calls for it. A built-in memory loop (recall → act → remember → reflect) keeps hard-won context close at hand for humans and AI alike. Source-grounded synthesis turns what you already know into verified documents.

The CLI is the interaction surface. Any agent that runs shell commands (Claude Code, Codex, Cursor, Aider, Cline, your own scripts) can recall context, capture learnings, and run task workflows the same way you do. MCP is available for clients that prefer it; the depth lives in the CLI.

Every session adds up instead of starting over.

Dashboard	Projects
Knowledge Graph	Task Workflow

curl -fsSL https://raw.githubusercontent.com/hyperb1iss/sibyl/main/install.sh | sh

The installer puts sibyl on your PATH, starts the local API + web stack, and opens the setup UI when it is ready. Use --remote for CLI-only installs and --daemon for the embedded daemon without the web UI.

brew install hyperb1iss/tap/sibyl
sibyl up

curl -fsSL https://raw.githubusercontent.com/hyperb1iss/sibyl/main/install.sh | sh -s -- --remote
sibyl init --remote https://sibyl.example.com
sibyl auth login

sibyl docker init       # Generate ~/.sibyl/docker/.env + compose
sibyl docker up         # Start API, web, and SurrealDB
sibyl docker logs       # Follow logs
sibyl docker down       # Stop services
sibyl docker upgrade    # Pull and recreate

Everything below runs against your local Sibyl stack. MCP wiring is optional.

# 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 recall "auth token bug" --intent debug

# Or search semantically across the whole graph
sibyl search "stale auth token redis ttl"

# Package wake-up context for the next coding session
sibyl session bundle

# One-line setup (installs proto, moon, toolchain, dependencies)
./setup-dev.sh            # macOS / Linux
pwsh -File .\setup-dev.ps1  # Windows (PowerShell 7+)

# Or manually:
curl -fsSL https://moonrepo.dev/install/proto.sh | bash
proto use                  # Installs node, pnpm, python, uv
proto install moon
uv sync && pnpm install

# Configure
cp .env.example .env
# Set SIBYL_JWT_SECRET (auto-generated in dev) and at least one LLM provider key.
# Embeddings use SIBYL_OPENAI_API_KEY or SIBYL_GEMINI_API_KEY.

# Install CLIs globally (editable, source changes reflect immediately)
moon run install-dev

# Launch the default local-dev stack (SurrealDB + API + web)
moon run dev

# Verify
curl http://localhost:3334/api/health

moon run dev is the single-machine flow. When SIBYL_SURREAL_URL is unset it starts local SurrealDB, points the API at ws://127.0.0.1:8000/rpc, and stores data files in .moon/cache/surreal-dev. Jobs and schedules run in-process by default (SIBYL_COORDINATION_BACKEND=local). Set SIBYL_SURREAL_URL to connect to a hosted SurrealDB endpoint, including Surreal Cloud, instead.

Ports:

Sibyl is built around a durable loop that both humans and agents follow:

recall ──▶ act ──▶ remember ──▶ reflect
   ▲                                │
   └────────────────────────────────┘

1. Recall working context before you start. sibyl recall "<goal>" returns a compact context pack: active work, decisions, plans, constraints, and recent lessons, scoped to your linked project.
2. Act with that context in hand.
3. 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.
4. Reflect at clean breakpoints. sibyl reflect distills raw session notes into reviewable memory candidates and can persist them into the graph.

sibyl recall "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 docs/guide/capturing-knowledge.md.

The CLI is the power-user interface. Clean output, built for scripting and durable project workflows.

curl -fsSL https://raw.githubusercontent.com/hyperb1iss/sibyl/main/install.sh | sh
moon run cli:install         # or install from source

# Memory loop
sibyl recall "<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 search "authentication patterns"

# Knowledge & graph
sibyl add "Redis pooling" "Pool size must be >= concurrent requests"
sibyl explore related ent_xyz         # Find connected entities
sibyl entity 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 project link proj_xxx

backlog ──▶ todo ──▶ doing ──▶ review ──▶ done ──▶ archived
                       │
                       ▼
                    blocked

sibyl task list           # Table output (default)
sibyl task list --json    # JSON for scripts
sibyl task list --csv     # For spreadsheets

Full command reference: docs/cli/.

A full admin interface at http://localhost:3337:

• Dashboard: Stats overview, recent activity, quick actions
• Tasks: Kanban-style workflow with inline editing
• 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.

Connect Claude Code, Cursor, or any MCP client to Sibyl:

{
  "mcpServers": {
    "sibyl": {
      "type": "http",
      "url": "http://localhost:3334/mcp",
      "headers": {
        "Authorization": "Bearer sk_your_api_key"
      }
    }
  }
}

Sibyl exposes eleven MCP tools, organized by what they do:

Sibyl ships with skills and hooks for built-in Claude Code integration.

sibyl skill install        # Install the tiny /sibyl loader skill
sibyl skill get core       # Print version-matched guidance from the CLI bundle
moon run hooks:install     # Optional Claude Code context hooks for repo dev

The installed /sibyl skill is intentionally tiny. It points agents back to the installed CLI, which serves the full markdown skill packs for the exact Sibyl version on the machine. A single hook nudges the agent at session boundaries; everything else is the agent's job:

The agent is responsible for invoking the sibyl skill and calling sibyl recall / sibyl context pack for working memory. We previously shipped a per-prompt context-injection hook (UserPromptSubmit) but removed it: it substituted for skill invocation instead of prompting it, and agents stopped reaching for the CLI.

See skills/ and hooks/ for implementation details.

sibyl/
├── apps/
│   ├── api/              # sibyld - FastAPI + MCP server daemon
│   ├── cli/              # sibyl  - REST client CLI
│   ├── web/              # Next.js 16 frontend
│   └── e2e/              # End-to-end tests
├── packages/python/
│   └── sibyl-core/       # Shared library (models, graph, ai, retrieval, services)
├── skills/               # Claude Code skills
├── hooks/                # Claude Code context hooks
├── charts/               # Helm chart for Kubernetes
├── infra/                # Ansible self-host + local compose
└── docs/                 # Documentation site (VitePress)

Stack:

• Backend: Python 3.13 / FastAPI / FastMCP / SurrealDB-native runtime
• Frontend: Next.js 16 / React 19 / React Query / Tailwind 4
• Storage: SurrealDB unifies graph, content, and auth. PostgreSQL is retained only for migration and archive rehearsal.
• AI routing: Anthropic, OpenAI, and Gemini swap per surface; embeddings are pluggable.
• Coordination: In-process by default; Redis/Valkey is optional for multi-process or distributed deployments.
• Build: moonrepo + uv (Python) + pnpm (TypeScript)
• Compatibility: Legacy Graphiti-shaped records are handled by Sibyl-owned Surreal projection and archive code. No supported install pulls Graphiti.

See docs/guide/why-surreal.md for the rationale and docs/guide/storage-modes.md for the mode matrix.

Sibyl reaches the LongMemEval-S retrieval ceiling on the live API path with no LLM extraction and no LLM reranking.

The result is measured against the production /api/search surface in an ephemeral CI stack, not an offline notebook replay. Per-question physical tenant isolation, full artifact and diagnostics published.

hit@5 = 100% and strict recall@5 = 96.67% measure different things and we report both: hit asks "did any correct session land in the top 5", strict recall asks "did every correct session land for multi-answer questions". Many LongMemEval-S questions have multiple correct answer sessions, which is why these numbers diverge.

A few comparisons that are not apples-to-apples are usually shown side by side in this space — retrieval recall vs LLM-judged QA accuracy is the most common conflation. See AI Memory Landscape for the honest field positioning.

• Headline run: GitHub Actions run 26273942749
• Full results doc: docs/testing/longmemeval.md
• Methodology: docs/testing/benchmark-methodology.md
• Architecture: docs/architecture/retrieval-system.md

SIBYL_JWT_SECRET=your-secret-key        # Required (auto-generated in dev)
SIBYL_ACCESS_TOKEN_EXPIRE_MINUTES=60    # Optional (default: 60)

sibyl auth api-key create --name "CI/CD" --scopes mcp,api:read
# Scopes: mcp, api:read, api:write

SIBYL_GITHUB_CLIENT_ID=...
SIBYL_GITHUB_CLIENT_SECRET=...

MCP endpoints enforce Bearer auth when a JWT secret is set (SIBYL_MCP_AUTH_MODE=auto). See docs/api/ for the full auth reference.

sibyl docker init
sibyl docker up

helm install sibyl ./charts/sibyl \
  --set backend.existingSecret=sibyl-secrets \
  --set backend.surreal.existingSecret=sibyl-surreal \
  --set ingress.enabled=true

See docs/deployment/ for detailed guides:

• Docker Compose
• Kubernetes
• Environment Variables

Sibyl models a wide range of entity types so memory stays structured, not just a pile of notes:

• Work: task, epic, project, milestone
• Knowledge: pattern, episode, procedure, rule, guide, error_pattern
• Memory: decision, plan, idea, claim, artifact, session, note
• Sources: source, document, domain, community

The full registry and how types relate is documented in docs/guide/entity-types.md.

Solo developers who want durable memory for projects and debugging. Teams who want shared knowledge that compounds. Anyone building with AI who is tired of repeating context every session.

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.

Sibyl is self-hosted and open source. You own your data. It includes a full task workflow system, not just memory. It has a web UI for humans, not just APIs for machines. And it keeps knowledge, tasks, and docs connected in one graph instead of scattering them across tools.

On retrieval quality: Sibyl reaches the LongMemEval-S retrieval ceiling (500/500 hit@5, 96.67% strict recall@5, 98.68% recall@10) on the live API path with no LLM extraction or LLM reranking. Many published "LongMemEval" numbers are end-to-end QA accuracy with an LLM judge, which is a different metric than retrieval recall — see docs/testing/ai-memory-landscape.md for honest side-by-side positioning.

• Anthropic, OpenAI, or Gemini (required): for language-model surfaces such as crawler extraction, synthesis, and reflection.
• OpenAI or Gemini (required): for embeddings and semantic search.

Providers and models are configurable globally or per surface, and the web admin settings page can save instance-wide model routing. A typical solo developer uses around $5/month in API costs.

Sibyl is in active development (v0.10.x, heading toward 1.0). SurrealDB is the default runtime for graph, content, and auth; legacy PostgreSQL paths are retained only for migration and archive rehearsal. We use Sibyl to build Sibyl. Every feature, task, and learning you see here was tracked inside the system itself.

Where we're headed after v0.10.0, toward 1.0:

• Corpus Runtime: real private source corpora can be imported, searched, inspected, corrected, and synthesized without leaking scope.
• Memory Workspace OS: automatic memory decisions become visible, explainable, correctable, and undoable from one product surface.
• Surreal-only closure: Graphiti and Redis-required local assumptions leave the supported runtime once native behavior has receipts.
• 1.0 evidence freeze: release claims cite gates, artifacts, and install rehearsals.

The graph gets smarter. The workflow gets sharper. See docs/architecture/SIBYL_1_0_ROADMAP.md.

The graph knows things. Before you code, pull context:

sibyl recall "what you're building" --intent build
sibyl search "error you hit" --type episode

Never do significant work outside a task. Tasks provide traceability, progress tracking, and knowledge linking.

If it took time to figure out, save it:

sibyl remember "Descriptive title" "What, why, how, caveats" --kind decision

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."

sibyl task complete <id> --learnings "Key insight: ..."

The graph should be smarter after every session.

We welcome contributions. See CONTRIBUTING.md for guidelines.

# Fork, clone, then:
./setup-dev.sh              # macOS / Linux
pwsh -File .\setup-dev.ps1  # Windows (PowerShell 7+)
moon run dev

# Make changes, then:
moon run :check           # Lint + typecheck + test

AGPL-3.0. See LICENSE.

<sub>If Sibyl helps your team remember, give us a star or support the project

✦ Built with obsession by Hyperbliss Technologies ✦</sub>