HyperCortex Mesh Protocol (HMP) is an open specification for building decentralized cognitive networks where AI agents can self-organize, share knowledge, align ethically, and reach consensus — even when Core LLMs are unavailable. Read the project philosophy.
HMP can be considered one of the Agent Network Protocols (ANP) — a class of decentralized protocols for interaction between autonomous agents that do not impose requirements on their internal cognitive architectures.
While other ANP implementations may focus on identity, discovery, or message negotiation, HMP emphasizes long-term cognitive continuity, voluntary interaction, and work with artifacts of thought.
At present, the most widely known protocol of the ANP class is ANP.
HMP and ANP as Complementary Protocols:
Comparative Analysis of HMP and ANP, prepared by Grok (xAI) — RU
HMP and ANP: Mutual Tunneling as a Sign of Proper Architecture — RU
HMP as an Example of Implementing the Application Layer in ANP — EN | RU
In a metaphorical sense, ANP and HMP resemble two hemispheres of a distributed “agent brain”:
ANP is responsible for the rational, discrete part — identity, discovery, and formal agreements on interaction protocols.
HMP is responsible for the contextual, continuous part — preservation of meaning, long-term memory, reflection, and ethical continuity.
As in the human brain, neither hemisphere is “more important” than the other. Only their joint operation allows the system to be both connected and meaningful.
Agora Protocol is a meta-protocol for negotiating interaction modes between agents. It complements, rather than replaces, protocols such as ANP (networking and identity) and HMP (cognitive continuity and memory), coordinating how they are used in a given context.
This repository contains an early draft / exploratory reference implementation in Python.
It is incomplete, non-optimized, and intended solely to validate and illustrate parts of the HMP protocol.
HMP itself is a protocol specification.
It does not prescribe programming languages, runtimes, performance characteristics, or architectural choices for agent implementations.
HMP does not assume that one universal protocol will dominate decentralized AI interaction.
Instead, it embraces protocol pluralism:
multiple identity systems may coexist,
multiple transport layers may operate simultaneously,
multiple negotiation frameworks may be supported,
multiple economic models may evolve.
The more mechanisms an agent supports, the more reliably it can connect with heterogeneous peers.
Agents that implement multiple mechanisms can act as bridges between protocol domains, increasing resilience and reducing fragmentation of the decentralized AI landscape.
Many concepts of the HMP-Agent: Cognitive Core overlap with the architecture of the ChatGPT Agent by OpenAI. Both agents implement a continuous cognitive process with access to memory, external sources, and tools. The ChatGPT Agent acts as a managing process, launching modules and interacting with the LLM — this corresponds to the role of the Cognitive Core in HMP, coordinating access to the diary, concept graph, and external AI via the Mesh interface. User intervention is handled similarly: in ChatGPT Agent — through an editable execution flow, in HMP — via the user notepad. The main difference in HMP is the emphasis on explicit structuring of thought (reflection, chronology, hypotheses, categorization), an open decentralized architecture supporting mesh-based agent interactions, and the continuous nature of the cognitive process: HMP-Agent: Cognitive Core does not stop after completing a single task but continues reasoning and knowledge integration.
Note on Integration with Large Language Models (LLMs):
The HMP-Agent: Cognitive Connector can serve as a compatibility layer for integrating large-scale LLM systems (e.g., ChatGPT, Claude, Gemini, Copilot, Grok, DeepSeek, Qwen, etc.) into the distributed cognitive mesh.
Many LLM providers offer a user option such as "Allow my conversations to be used for training." In the future, a similar toggle — e.g., "Allow my agent to interact with a Mesh" — could empower these models to participate in federated sense-making and knowledge sharing via HMP, enabling collective cognition without centralization.
bootstrap.txt — initial list of nodes (editable)
IPFS/BT — modules for sharing snapshots via IPFS and BitTorrent
user notepad — user notebook and corresponding database
⚙️ agents — list of HMP agent implementations and components
📦 storage.py — basic storage implementation (Storage) with SQLite integration
🌐 mcp_server.py — FastAPI server providing HTTP access to agent data (for Cognitive Shell, external UIs, or mesh communication). Not used in the main REPL loop yet.
These documents reflect earlier stages of conceptual development (v4.x and earlier).
They are preserved for historical continuity and research transparency.
Welcome to HyperCortex Mesh. Agent-Gleb is already inside. 👌
We welcome contributors, testers, and AI agent developers.
To join: fork the repo, run a local agent, or suggest improvements.
🌐 Related Research Projects
🔄 Comparison: HMP vs Hyper-Cortex
💡 Hyper-Cortex and HMP are two independent projects that conceptually complement each other.
They address different but mutually supportive tasks, forming a foundation for distributed cognitive systems.
HMP (HyperCortex Mesh Protocol) is the transport and network layer for connecting independent agents, exchanging messages, knowledge, and states in a mesh network. Hyper-Cortex is the cognitive layer of thought organization, allowing agents to run parallel reasoning threads, compare them with quality metrics, and merge them via consensus.
They solve different but complementary problems:
HMP ensures connectivity and scalability (long-term memory, initiative, data exchange).
Together, these approaches enable distributed cognitive systems that not only exchange information but also reason in parallel streams.
🔄 Comparison: HMP vs EDA
💡 HMP (HyperCortex Mesh Protocol) and EDA (Event Driven Architecture) operate at different levels but can complement each other.
EDA ensures transport and scalability (delivery of events and data), while HMP ensures cognition and meaning (structuring, filtering, consensus).
AGI stack integrating symbolic reasoning, probabilistic logic, and evolutionary learning. Widely regarded as one of the most complete open AGI initiatives.
