What is memforks?

MemForks applies version control concepts, similar to how Git manages code history, to the memory that AI agents build up as they work. Today most AI agent frameworks store memory as a flat list of things the agent has recorded, with no way to branch off and explore a risky line of reasoning without affecting the main memory, no way to run multiple agents in parallel with isolated state, and no structured way to reconcile conflicting memories when agents merge their work back together. MemForks is designed to solve those problems. The core model looks like a Git commit graph: memory has a main branch, you can create forks to explore alternative hypotheses, parallel agents each work on their own branch, and when work is ready to merge, a resolution process settles conflicts using typed rules rather than simply letting the last write overwrite everything. Every conclusion can be traced back to the specific commit that introduced it, making the agent's reasoning auditable. The system is built on top of a blockchain called Sui, which provides cryptographic anchoring and finality for merges. Memory blobs are stored using two related projects: MemWal for retrieval and Walrus for underlying blob storage, with SEAL handling encryption. The merge resolution step involves a group of attestors who sign off on how a conflict is resolved before the result is finalized on-chain. At the time of writing, the protocol is deployed on Sui testnet and the repository includes links to live on-chain objects and transactions showing a completed branch-merge cycle. However, the full open-source code release, including the TypeScript SDK, command-line tool, framework adapters for LangGraph and Vercel AI SDK, and a visual graph explorer, has not yet shipped. The README describes these as coming soon and asks visitors to watch the repository for the release notification. The project is Apache 2.0 licensed.