The lifelong-learning agent runtime

Open-source. MIT-licensed. Five-layer composed memory. Reflexion at session-end. Voyager-style skill compilation. Multi-agent debate. Built in Frankfurt.

Thoth — Egyptian god of writing, wisdom, and the moon. Patron of scribes. Recorder of words. Weighed each soul against a feather and wrote the verdict.
npx @thoth-runtime/cli init my-thoth

Memory composes in five layers

Each layer has its own writer, its own cadence, its own kind of truth. None can collapse into the others without losing what makes a mind a mind.

L5

Reflection

Reflexion at session-end with four fan-out writers (memory, skill drafts, persona observations, Honcho updates). Founder-gated; nothing auto-applies.

L4

Procedural

Skills, persona files, hard rules. Voyager-style skill compilation: patterns recur, become candidates, you approve with a single reaction.

L3

Episodic

Cosine recall over local Xenova embeddings (384-dim, ~50ms/encode, $0/encode). Top-3 related episodes inject as context on every new thread.

L2

Identity

Theory-of-mind on each peer via Honcho. Workspace · Peer · Session · Message — relational facts that evolve across all conversations.

L1

Working

Current session state. Auto-managed by the model itself via the CLAUDE.md file pattern.

What ships in v0.5

A composed runtime, not a library. Opinions on layering, reflection, skill evolution.

Reflexion at session-end

Every session ends with a reflection step. JSON output fans out to four writers — memory, skill drafts, persona observations, Honcho. Founder-gated; nothing auto-applies.

Voyager-style skills

Patterns that recur become candidate skills. Approve with a single ✅ reaction. Skill genealogy + RL refinement land in v0.6.

Reactions as training signals

✅❌🧠🗑️👤 on bot replies feed reflection, skill drafts, identity model updates. The chat surface is the labelling surface.

Multi-agent council

Convene a party of personas to debate a question. v0.5 homogeneous; v0.6 mixes models — Mary on GPT-5, John on Claude Opus, Winston on Gemini.

Mythic dashboard

Tree of Life, Akashic Records, Council, Firmament, 3D Cosmos. Every internal state inspectable. Every decision auditable.

Slack-native

The chat surface IS the agent surface. Threaded sessions, magic commands, ephemeral debug. Discord ships Q3.

vs. agent libraries

LangChain, Letta, Mem0, CrewAI are libraries. Thoth is a runtime.

LangChain / CrewAI

Stateless by default. You assemble the architecture per project.

Thoth

Stateful by design. Architecture ships with opinions on layering, reflection, skill evolution.

Letta / Mem0

Memory libraries. You wire the rest yourself.

Thoth

Five composed layers + reflection + skills + multi-agent + chat-native UX as one organism.

Devin / Cursor / Replit Agents

Closed runtimes. Architecture lives inside one company's walls.

Thoth

Open-source MIT runtime. Closed pieces are Cloud + marketplace; the architecture is yours.

Anthropic Claude Skills / OpenAI Agents SDK

Vendor-specific. Locked to one provider's models.

Thoth

Multi-model. Subscription-billing-friendly via local Claude CLI delegation. Mixes providers in v0.6 party.

Open core. No surprises.

Run the open-source runtime yourself for free, forever. Or let us host it.

Self-host

€0/forever
  • MIT-licensed runtime
  • All five memory layers
  • Slack transport
  • Local Claude CLI billing
  • Community Discord support

Free Cloud

€0/month
  • 1,000 episodes/mo
  • 1 workspace, 1 seat
  • 30-day retention
  • Community Discord
  • EU residency

Enterprise

from €1,500/month
  • Unlimited everything
  • SSO + audit log
  • On-prem option
  • Dedicated support
  • Custom DPA + MSA

Cloud opens for paying customers in Q3 2026. Sign up to be notified.

Ready?

The 5-layer memory diagram is in your terminal in 60 seconds.