Ditto’s Two-Subnet Architecture to Combine Persistent Memory With High-Speed Inference

Novelty Search E082 featured the Ditto team unveiling the latest progress on SN118 and announcing Ditto Cloud, a second subnet built with Beyond Finance for fast verifiable inference.

Ditto (SN118) is a decentralized memory layer that lets AI agents share persistent context across Claude, Codex, Hermes, Cursor, and any MCP-compatible interface.

Ditto Cloud targets frontier-level inference speeds of 14,000 to 20,000 tokens per second inside a Trusted Execution Environment (TEE), enabling large-scale simulation and background reasoning.

Together, the two subnets form a unified memory-and-inference stack that keeps AI agents persistent, context-aware, and responsive in real time.

Technical Highlights From Novelty Search E082

The conversation moved through Ditto’s current architecture, the Cloud subnet announcement, the dreaming pipeline, and the new Bittensor mining primitives making all of it possible.

1. Ditto (SN118) lets multiple agents share the same memory graph: Claude, Codex, Hermes, Cursor, and any other MCP-compatible interface can plug into the same underlying context, meaning switching platforms no longer means losing history.

2. The real user demo landed clean: Switching from OpenClaude to Hermes with Ditto let the Hermes agent inherit everything the OpenClaude agent knew, described on stage as “bloop, I know Kung Fu.” Zero context loss.

3. Storage architecture is production-grade: Hetzner instances with Postgres for the graph database, object storage for larger artifacts, and Hippius (SN75) buckets for users who want their objects stored on Bittensor infrastructure.

4. Full encryption is next on the roadmap: Every piece of data stored on Ditto servers will be encrypted, with the team already exploring TEE-based inference for sensitive pipelines.

5. The dreaming pipeline runs background curation while users are offline: Every prompt and response triggers a dreaming cycle that clusters memories into subjects, merges related ones, generates rolling summaries, and organizes files into folders.

6. The design philosophy is vacuum-first: Suck up as much data as possible without worrying about cleaning it. The dreaming pipeline handles refinement, compression, and hierarchical clustering automatically.

7. Auto-detected workflows turn into automations: When Ditto notices a repeating pattern in how a user works, it extracts the pattern and suggests it as an automation, with generative UI templates for setup.

8. Agent-native signup replaces API key fumbling: Secure Shell (SSH) into a new machine, run the universal prompt, and the agent creates a claimable account. The user clicks one link to merge that agent into their main workspace.

9. DittoBench uses synthetic users every run: A random fake user gets generated each cycle for anti-gaming, with needle-in-haystack retrieval on datasets that exceed model context windows by multiple orders of magnitude.

10. Ditto Cloud will target FPGA and analog computing: Long-term goal is competing with Talaas, Canada at 14,000 to 20,000 tokens per second.

Analog computing bakes model weights into resistors and ohms directly on a chip, potentially fitting a DeepSeek R1-sized model onto a single Intel Xeon-sized board.

The Layer Underneath Your Agent

Ditto’s vision is that memory is the foundation of the AI agent experience, with storage, dreaming pipelines, cross-platform sharing, and Ditto Cloud all built to support it.

Together, the two subnets combine persistent memory with fast verifiable inference, making AI agents feel more like operating systems than isolated chat tools.

The architecture is a bet that the next wave of value in the agent economy will come from context that persists across interfaces and inference that runs at real-time speeds.

For users tired of losing context or juggling multiple AI tools, Ditto is building the infrastructure that makes those problems disappear.