llm-sync

CRDT and vector clock primitives for distributed LLM agent state synchronization.

Coordinate shared state across multiple LLM agents without a central coordinator. Merge concurrent updates safely using conflict-free replicated data types.

What's inside

VectorClock — happens-before ordering for distributed agent events
GCounter / PNCounter — grow-only and increment/decrement CRDTs for agent metrics
LWWRegister — last-write-wins register for agent configuration and shared flags
ORSet — observed-remove set for shared agent capability registries
StateMerge — merge any two CRDT states deterministically

Use cases

Shared agent configuration that multiple instances can update concurrently
Distributed rate-limit counters across an agent fleet
Consensus-free agent state replication over NATS or Redis pubsub
Conflict-free merging of agent memory segments

Quick start

use llm_sync::{VectorClock, LWWRegister};

let mut clock_a = VectorClock::new("agent-a");
let mut clock_b = VectorClock::new("agent-b");

clock_a.increment();
clock_b.increment();

// Merge — safe under concurrent updates
clock_a.merge(&clock_b);
println!("Merged clock: {:?}", clock_a);

Add to your project

[dependencies]

llm-sync = { git = "https://github.com/Mattbusel/llm-sync" }

Test coverage