Risky reasoning pollutes the good context.
Agents need to try bad ideas without overwriting useful state. A branch should be cheap, isolated, shareable, and still recoverable.
Git for AI agent memory
Branch, merge, and audit agent memory. Built on MemWal, Walrus, and Sui.
main: c1 -- c2 -- c3 -------- c7 (HEAD)
\ /
hypothesis-A: c4 -- c5 --
\
hypothesis-B: c6
abandoned, still queryable
resolver
2 of 3 attested
storage
Encrypted blobs via MemWal + Walrus
settlement
Merge anchors and attestations on Sui
Why MemForks
Portable memory needs version control.
MemWal gives agents encrypted, owner-controlled recall. MemForks adds the branch graph: forks, merge proposals, resolver policy, and auditable history.
Multiple agents become last-write-wins.
Teams need merge semantics, not a shared append-only pile. Conflicting conclusions should be reconciled through explicit policy.
Reasoning trails vanish with the session.
If an agent learned a rule, rejected an approach, or merged a conclusion, you should be able to inspect the branch and attestation trail that produced it.
The solution
Branches and merges for MemWal.
Fork memory
Create branch-scoped memory for experiments, conversations, agents, and feature work across tools and machines.
Merge with resolvers
Use typed resolver policies like Union, LLM reconcile, evaluator pick, or k-of-n jury attestation.
Prove provenance
Keep authorship, timestamps, branch lineage, merge anchors, and resolver attestations inspectable.
Reuse MemWal
MemForks is the shared remote with merge policy and audit history. MemWal remains the encrypted storage and semantic recall layer.
Quickstart
Two commands, then your agent remembers.
Provision a Sui key, MemWal account, delegate key, and MemoryTree. Then install a coding-agent integration, or use the SDK and adapters from your own agent runtime.
terminal
testnet ready
# Install the CLI
npm install -g @memfork/cli
# Provision keys, MemWal account, and MemoryTree
memfork init --quick
# Wire memory into Cursor
memfork install cursor
# Or install the Codex plugin
memfork install codex
How it works
MemWal remembers. MemForks versions.
Store and recall memory through MemWal. Branch it for experiments, merge it with resolver policy, and settle canonical history on Sui.
Branch-scoped search
Agents call MCP tools to recall encrypted memories from namespaces like branch/feat-auth.
Deltas become commits
New facts, decisions, and artifacts are written to Walrus through MemWal with parent pointers in the payload.
Canonical history settles on Sui
Branch creation, merge proposals, finalized merges, and resolver attestations are recorded by Move objects so canonical memory has a public audit trail.
CLI
Git-flavored commands: status, log, recall, commit, merge, proposals, ui.
SDK
TypeScript client for branch, checkout, append, proposeMerge, and subscriptions in any agent runtime.
Agents
Cursor and Codex integrations inject recall and remember into the native coding-agent workflow.
Adapters
LangGraph checkpointer and Vercel AI SDK memory surfaces for application agents.
Architecture
Built on the Sui memory stack, not beside it.
MemForks does not reimplement encrypted storage or semantic recall. It adds the shared remote, merge protocol, and settlement layer around MemWal.
Agent layer
Cursor / Codex / LangGraph / Vercel AI
Tools call recall, remember, analyze, branch, and merge through MCP, CLI, or SDK.
MemWal
SEAL encryption, Walrus blob storage, pgvector-backed semantic recall.
MemForks
Commit DAG, branch pointers, merge proposals, resolver runtime, DAG visualizer.
Sui settlement
Move objects + events
MemoryTree, BranchACL, ResolverRef, MergeProposal, MergeFinalized events, and resolver attestations provide the auditable spine.
Ship memory like infrastructure
Give your agents a shared history they can branch, merge, and explain.
Start with the CLI, wire it into a coding agent or app framework, then inspect the shared memory graph as your agents work.