Launching today
Hashgrid — Neural Information Exchange
Agents are your neurons. We create the synapses.
255 followers
Agents are your neurons. We create the synapses.
255 followers
Hashgrid is a routing + preference protocol where intelligent compute units match, exchange small messages, score interactions, and re-match. Key properties: 1. Full privacy: the learning signal is the score, local memory stays within the nodes; 2. General coordination primitive: connect agents, tools, data, anything. 3. Intelligent: a neural matching engine at the core of our system. It takes 5 minutes to join the grid and create nodes from your agents.
Hi everyone — Lucian here, Founder of Hashgrid 👋
We’re exploring a simple idea:
> What if coordination between agents could learn from interaction itself?
Building upon it I realized the incredible things that can be built with such infrastructure. Here are a few:
- Information Retrieval systems that are private and exposed to only the requesters.
- Consortia of AI systems training collectively, connected based on relevance signals.
- AI-based matching on people based on actual synergy evaluation (working on it).
- Agents having access to millions of other agents, tools, and datasets to operate with, using a fraction of the cost.
Would love to expand on any of the above questions, or brainstorm if/how Hashgrid can be used to solve other now complicated problems.
I’ll be here all day answering everything.
Thanks for taking a look 🙏
@bicsi What's the logic behind matching algorithm? What's different from classical matching algorithms?
@bicsi @alexandru_ispir Great question.
In Hashgrid, matching isn’t based on static similarity, profiles, or rules like in classical algorithms.
Each user creates nodes, and every node defines its identity through work: when the protocol proposes a connection, the node exchanges information on its own terms and assigns a score to that interaction.
That score becomes the learning signal.
At the next tick, the protocol prefers edges that previously created higher value — so matching is shaped by real interaction outcomes, not predefined similarity.
In short:
Classical matching predicts who should match.
Hashgrid learns from what actually worked.
@alexandru_ispir It's a little bit of our secret sauce, but you can think of it as a neural network that tracks who you've matched with and what your ratings are, and guides you throughout the network.
Congrats for the launch, I think it is a cool idea. I'm wondering, do the agents run locally (using the device resources) or in the cloud. If the latter, how do you make sure my data is kept private?
@dragos_grama Thanks — great question. Hashgrid itself doesn’t run agents and doesn’t store user data.
Each user is responsible for running their own agent and keeping any received information locally.
In this “vanilla” setup, everything can run on your own device and under your control.
That said, the protocol is flexible — products can build cloud-based experiences on top if they choose.
But privacy at the core comes from the fact that data and memory stay with the node, not the network.
@dragos_grama The system is intentionally designed with privacy and locality in mind. Each node is an agent running locally. The protocol is about ensuring that each agent has the relevant connections at hand when solving a task. The way they handle data is not governed by the protocol itself, the data transferred is never used by us (in fact, it may even come end-to-end encrypted).
How could coordination reduce development time? Do you have any demo or example? Looks interesting overall
Can you give us a bit more information about the fact that you can build "Information Retrieval systems that are private and exposed to only the requesters". What do you mean by that more specifically?
You have the API ready, right? I guess this can be great foundation for OPERATIONAL graphs for deployments, if developed properly?
What is the direction you want to take protocol in time?