Hyperbliss Technologies just dropped Sibyl on Hacker News—a self-hosted, cross-agent memory system that gives all your AI coding tools one shared brain to work with. The pitch is straightforward: instead of each agent starting from scratch every session, Sibyl maintains a compounding knowledge graph of decisions, patterns, gotchas, and conventions across your entire codebase history. It's CLI-first, open source under Apache-2.0, and works with Claude Code, Codex, Cursor, or any agent you roll yourself.

The Memory Loop at the Core

Sibyl builds around a durable recall → act → remember → reflect cycle that both humans and agents follow. Before starting work, you run sibyl recall to pull a compact context pack covering active tasks, recent decisions, and relevant lessons scoped to your project. After learning something durable—like why authentication broke or which Redis TTL caused a production incident—you hit sibyl remember with the details so future sessions don't rediscover it. The sibyl reflect command then distills raw session notes into reviewable memory candidates that can be promoted into the graph after automatic grading.

CLI-First, Agent-Native Design

The interaction surface is the shell—if your agent can run a command, it already speaks Sibyl. No SDK adoption required, no lock-in to a specific MCP server configuration. The skills system makes this concrete: sibyl skill install drops a loader into Claude Code's ~/.claude/skills, Codex's ~/.codex/skills, and the generic ~/.agents/ convention simultaneously. The installed skill is deliberately tiny—it just points back at the CLI. Real guidance lives in version-matched packs served on demand via sibyl skill get, so upgrading Sibyl keeps all agent guidance in sync automatically.

Technical Architecture Under the Hood

The backend runs Python 3.13 with FastAPI and FastMCP, while the frontend uses Next.js 16 and React 19 atop Tailwind CSS 4. SurrealDB handles graph storage, content, and authentication in one runtime, with Redis or Valkey optional for multi-process deployments. On retrieval quality, Sibyl claims it hits 100% hit@5 on LongMemEval-S through the live /api/search path without LLM extraction or reranking—though the team notes many published benchmarks measure end-to-end QA accuracy rather than raw retrieval recall, making direct comparisons tricky.

MCP Integration and Web UI

For clients that prefer it, Sibyl exposes eleven MCP tools covering search, context compilation, graph exploration, knowledge creation, synthesis drafting, and task management. The web UI at localhost:3337 provides a full admin interface with dashboard stats, Kanban-style task workflow, an interactive force-directed graph visualization, semantic search with filters, memory workspace, source configuration, and settings for organizations, teams, API keys, and LLM routing.

What Comes Next

Sibyl hit 1.0 status as of this launch—the team uses it to build itself, which is either the best or worst QA strategy depending on your tolerance for recursion. Post-1.0 priorities include making automatic memory decisions visible and correctable from one surface, propagating edits and redactions forward into future recall and synthesis, and expanding source ingestion beyond mailbox-style archives into broader corpora with scope and provenance tracking.

Key Takeaways

  • One knowledge graph shared across every AI coding tool you run—no more starting fresh each session
  • CLI-first design with optional MCP integration; agents speak Sibyl if they can execute shell commands
  • Skills system auto-installs cross-platform (Claude Code, Codex, generic) and stays version-matched via on-demand packs
  • Self-hosted means your context, decisions, and accumulated knowledge stay yours—no vendor lock-in or profile harvesting

The Bottom Line

Sibyl scratches a real itch for developers running multiple AI agents in parallel—if you've ever watched two different coding assistants make opposite mistakes because neither remembered what the other figured out, you already know why this matters. Whether the memory loop discipline actually gets followed depends on whether it stays frictionless enough that using Sibyl feels faster than not using it.