Install
$ agentstack add mcp-hmg-ai-hmg-public Open-source listing, not yet scanned by AgentStack. Follow the source repository for install instructions.
Security review
⚠ Flagged1 finding(s); flagged for manual review. · v0.1.0 How review works →
- • Prompt-injection patterns
- • Secret / credential exfiltration
- • Dangerous shell & filesystem operations
- • Untrusted network calls
- • Known-malicious package signatures
- high Pipes remote content directly into a shell (remote code execution).
What it can access
- ● Network access Used
- ✓ Filesystem access No
- ✓ Shell / process execution No
- ✓ Environment & secrets No
- ✓ Dynamic code execution No
From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.
Reliability & compatibility
Declared compatibility
Compatibility is declared by the source manifest. End-to-end runtime verification is coming, see below.
We're building live execution health for every listing: tool-call success rate, median latency, uptime, and last-checked timestamps, measured, not self-reported. It isn't live yet, so we don't show numbers we can't stand behind.
How agent discovery & health will work →About
HMG — Agent Memory That Actually Works
Holographic Memory Graph — branch-aware, auditable, governable long-term memory for AI coding agents. Every session forgets. HMG makes sure the next one remembers.
🌐 Docs · 📦 Releases · 🏠 Website · 🚀 Quick Start · ✨ What's New
The Problem
Your AI coding agent is brilliant — and amnesiac.
Every new session starts with a blank slate. It re-reads the same files, re-asks the same questions, re-litigates decisions you already settled. "Why are we using Postgres?" gets answered six times across six sessions. A bug fix from last week? Forgotten. The API key it shouldn't commit? Committed again.
The cost isn't just keystrokes — it's trust. You can't delegate to an agent that keeps losing the thread.
Why HMG?
HMG gives an agent the one thing it's missing: a durable, queryable, governable memory.
| Without HMG | With HMG | |---|---| | Re-reads the whole repo to relearn context | Wakes up with a one-shot brief: status, decisions, risks, next steps | | Re-argues settled architecture questions | Knows why a choice was made — and what was rejected | | Silently overwrites the old answer when it's wrong | Keeps full correction history; the new answer is canonical, the old is never lost | | Commits secrets and sensitive context | Governance quarantines, seals, or derives safe lessons | | Forgets your branch the moment you switch | Memories are branch-scoped: tenant → workspace → repository → branch |
One tool call returns complete context. No prompt-stuffing. No context window limits. No forgetting.
Quick Start
# Install (Linux / macOS)
curl -fsSL https://github.com/HMG-AI/HMG-public/releases/latest/download/install.sh | sh
# Initialize + start the daemon
hmg init -g
hmg daemon start
# Install (Windows PowerShell)
irm https://github.com/HMG-AI/HMG-public/releases/latest/download/install.ps1 | iex
# Initialize + start the daemon
hmg init -g
hmg daemon start
Store your first memory and pull it back — same works for any HTTP client:
# Store a decision
curl -s -X POST http://127.0.0.1:7654/api/memorize \
-H 'Content-Type: application/json' \
-d '{"content": "Decision: use Rust for performance-critical paths"}'
# Recall it later
curl -s http://127.0.0.1:7654/api/recall \
-d '{"query": "what language for perf?"}'
# Verify your setup is healthy
hmg doctor
> Community Edition is free — 100,000 memory atoms, all core features, no license key. No account, no telemetry, no cloud. Data stays local.
The Agent Memory Loop
capture → canonicalize → recall → correct → govern → consolidate → brief next session
Every session becomes context for the next. Decisions persist. Corrections leave a traceable history. Governance controls what gets remembered — and what gets forgotten.
# Store a decision with rationale
memory_memorize(content="Use Rust for performance-critical paths", source="ADR-001")
# Next session: the agent wakes up with full context
memory_agent_brief(domain_pack_id="software-engineering")
# → scope, current status, prior decisions, unresolved risks, next steps
# Correct stale knowledge — the old version is kept in history
memory_correct(target_atom="01ABC...", action="replace", new_content="Use Rust + WASM")
# Govern sensitive data so it never leaks into recall
memory_govern(target_atom="01DEF...", action="seal", reason="contains API key")
Core Capabilities
Not a vector database. Not a key-value store. HMG models memory as a typed graph — every atom carries polarity, epistemic state, governance tags, and a correction lineage.
- 🧲 One-Shot Recall — A single MCP call returns complete context: current status, prior decisions, unresolved risks, and next steps. No re-reading, no re-asking.
- 🌿 Branch-Aware Scope —
tenant → workspace → repository → branch. Memories stay isolated per branch and become traceable on merge — yourfeature/authwork doesn't bleed intofeature/payments. - 🛡️ Governance Control Plane — Quarantine, seal, tombstone, or derive lessons. Policy tags and audit context flow through every operation. Normal recall skips governed content; audit recall still sees it.
- 🏠 Local-First — Embedded storage, zero external dependencies. Your memory lives on your machine, works offline, and never phones home.
- 🔌 47 MCP Tools — Memorize, recall, correct, govern, history, handoff, agent-brief, MemoryQL, observation capture, vault, panorama, and graph-health workflows — all over a single standard interface.
- 📜 Open Protocol — [
hmg-protocol](protocol/) (Apache-2.0) is a standalone crate defining types and serialization for every memory operation. Implement it, verify against it, build on it.
What's New in v1.6
- 🤝 Cross-Client Reliability — Spec-correct MCP
ping+ capability probes, and a transparent daemon-proxy fallback when a direct server hits a store-lock collision. Works smoothly with strict clients like opencode and the Hermes gateway. - 🧬 Agent-Native Extraction — A dedicated
DomainLensregistry lets recall and anchoring lenses evolve independently of the coarser extraction profile. Caller-supplied entities and relations thread straight through the memorize pipeline. - 🎯 Answer vs Context — Recall now flags the top-ranked answer cluster as
primary, distinct from same-scope supporting context — so agents surface the answer first instead of drowning in context. - 🔐 Two-Tier Source License — HMG is proprietary commercial software. The 19 core/product crates ship under the HMG Proprietary Source License; the 3 ecosystem-edge crates (
hmg-domain-packs,hmg-sdk,hmg-evals) stay Apache-2.0.
See [CHANGELOG.md](CHANGELOG.md) and Releases for the full history.
Editions
| | Community (Free) | Developer ($9/mo) | Enterprise | |---|---|---|---| | Memory atoms | 100,000 | Unlimited | Unlimited | | Semantic search | ✓ | ✓ | ✓ | | One-Shot Recall | ✓ | ✓ | ✓ | | Correction + Governance | ✓ | ✓ | ✓ | | Observation capture | ✓ | ✓ | ✓ | | Domain pack | software-engineering | software-engineering | All | | Consolidation | — | ✓ (auto) | ✓ (full) | | Intelligent lifecycle | — | ✓ | ✓ | | SSO / RBAC | — | — | ✓ |
One binary, runtime edition detection. No reinstall, no migration. Start free on Community and upgrade anytime with hmg license apply .
Agent Adapters
HMG works with the agent you already use. Wire any of the 19 supported adapters — each is doctor-verifiable:
| Agent | Init command | |-------|--------------| | pi | hmg init --agent pi | | Codex CLI | hmg init --agent codex | | Cursor | hmg init --agent cursor | | Claude Code | hmg init --agent claude | | VS Code | hmg init --agent vscode | | opencode | hmg init --agent opencode | | Continue | hmg init --agent continue | | Cline | hmg init --agent cline | | Roo Code | hmg init --agent roo-code | | Windsurf | hmg init --agent windsurf | | Zed | hmg init --agent zed | | Aider | hmg init --agent aider | | OpenHands | hmg init --agent openhands | | Goose | hmg init --agent goose | | Gemini CLI | hmg init --agent gemini-cli | | Qwen Code | hmg init --agent qwen-code | | OpenClaw | hmg init --agent openclaw | | Hermes | hmg init --agent hermes | | Any MCP client | hmg init --agent generic-mcp |
Run hmg doctor to verify every adapter's lifecycle wiring, store path, and MCP readiness, or hmg integrations detect to see which agents are installed in your environment.
SDKs
# Python
pip install hmg-sdk
import hmg
client = hmg.HmgClient()
client.memorize("key decision: use Rust for perf")
# TypeScript
npm install @hmg_ai/sdk-ts
import { HmgClient } from "@hmg_ai/sdk-ts";
const client = new HmgClient();
await client.memorize({ content: "decision noted" });
What's in This Repository
This repository holds the public artifacts for HMG — everything you need to install, integrate, verify, and implement compatible protocols. The memory intelligence engine itself is proprietary.
| Path | What it is | |------|-----------| | [protocol/](protocol/) | hmg-protocol (Apache-2.0) — standalone crate defining all memory types & serialization | | [mcp/schemas/](mcp/schemas/) | MCP tool definitions for all 47 public tools | | [openapi/](openapi/) | HTTP API specification | | [sdk-python/](sdk-python/) · [sdk-ts/](sdk-ts/) | Official Python and TypeScript SDKs | | [certification/](certification/) | Conformance tests for compatible implementations | | [docs/](docs/) | Multilingual documentation (10 languages) + live site | | [scripts/](scripts/) | Installers — install.sh, install.ps1 |
Documentation
| | | |---|---| | [Getting Started](docs/md/getting-started.md) | Install → first memory in 5 minutes | | [Concepts](docs/md/concepts.md) | Memory atoms, correction, governance, scope | | [API Reference](docs/md/api-reference.md) | MCP tools and HTTP endpoints | | [Architecture](docs/md/architecture.md) | System overview | | [FAQ](docs/md/faq.md) | Common questions | | [Upgrade Guide](docs/md/upgrade.md) | Version-to-version migration notes |
Links
- Website: hmg1ai.com
- Docs: hmg-ai.github.io/HMG-public
- Releases: GitHub Releases
- Issues: GitHub Issues
- Security: GitHub Security
- Contributing: [CONTRIBUTING.md](CONTRIBUTING.md)
License
HMG uses a two-tier license model:
- Protocol & SDK artifacts (this repo): [Apache-2.0](LICENSE) — open, implementable, verifiable
- Community Edition binary: [LICENSE-COMMUNITY](LICENSE-COMMUNITY) — free to use, no redistribution
- Memory intelligence engine: proprietary (HMG Proprietary Source License)
Copyright © 武汉凡尘合创科技有限公司 / HMG AI.
Source & license
This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.
- Author: HMG-AI
- Source: HMG-AI/HMG-public
- License: Apache-2.0
- Homepage: https://hmg1ai.com/
Install and usage instructions live in the source repository linked above.
Reviews
No reviews yet, be the first.
Write a review
Versions
- v0.1.0 Imported from the upstream source.