Hy Memory 2.0 and MemHub Guide: Share Memory Across Codex and Claude Code

Switching between coding agents often means repeating the same project rules: use pnpm, do not edit generated migrations, run a particular test suite, and preserve a specific API contract. Hy Memory 2.0 and MemHub, released by Tencent Hunyuan on October 16, 2026, address that problem with a memory layer that is separate from any one model or client. With explicit permission, MemHub can turn Codex and Claude Code sessions into local memory and expose relevant context back to supported tools through MCP.

This guide focuses on the parts you can verify today: which layer to choose, the release and platform checks to perform before installing MemHub, what its Codex and Claude Code integrations change, and where its local-first privacy boundary ends. Sources and release assets were checked on October 20, 2026.

Key points before you install

• MemHub is the user-facing, local-first application for sharing approved memory across multiple AI tools.
• Hy Memory SDK is the Python memory engine for developers building chat, productivity, or coding agents.
• Hy-Memory-30B-A3B is an open memory-writing model. You do not need to self-host the 30B model just to evaluate MemHub.
• The public v1.0.4 assets include a macOS Apple Silicon DMG and Linux x64/arm64 runtimes. No Windows or Intel Mac installer was present in the checked release.
• Codex and Claude Code do not automatically receive every saved memory in every prompt. A managed Skill tells the model when memory may help, and the model actively calls MCP tools.

The model reads and writes memory, the SDK manages storage and retrieval, and MemHub connects memory to AI tools. Source: Tencent Hunyuan Hy Memory 2.0 page, checked October 20, 2026.

Choose the right layer

Layer	Best fit	You manage
MemHub	People who want Codex, Claude Code, Cursor, and other tools to reuse approved context	Connected apps, sync and recall grants, model provider, and local data
Hy Memory SDK	Developers adding durable memory to an agent or internal service	LLM and embedding endpoints, stores, scenes, modes, and retention
30B-A3B model	Teams replacing the memory writer in an existing framework or running research	Inference infrastructure, evaluation, and safety policy

For an individual developer who mainly wants project conventions to survive a tool change, MemHub is the practical starting point. Product teams building their own agent should evaluate the SDK instead. Its chat, working, and coding scenes organize personal conversation, task facts and reusable experience, and code decisions with version-evolution context in different ways.

Four checks to make before installation

1. Confirm the assets, not only the support target

The documentation describes a target matrix that includes macOS 13+ and Linux x64/arm64, but an installable platform must also have a real artifact in GitHub Releases and an entry in release-manifest.json. The checked v1.0.4 release contained a notarized Apple Silicon DMG, a macOS arm64 runtime, and Linux x64 and arm64 runtimes. It did not contain Windows or macOS x64 packages.

2. Keep the Bootstrap CLI and runtime on the same version

The npm registry listed memhub-cli 1.0.5 as latest, while the public GitHub release was v1.0.4. MemHub's installer is designed to verify an exact-version runtime manifest, so do not force an unmatched combination merely because one number is newer. For a cautious v1.0.4 evaluation, inspect the versions and pin the CLI to the release that actually has matching artifacts:

node -v
npm view memhub-cli versions --json
npm install -g memhub-cli@1.0.4
memhub install

The first memhub install is a planning step. It checks the release manifest, operating system, architecture, size, SHA-256 hashes, commit identity, and protocol compatibility before applying the installation. Continue only after reviewing the plan, using memhub install --yes. The published CLI requires Node.js 24 or newer.

3. Understand where model processing happens

MemHub is local-first, not necessarily fully offline. If you select a third-party model provider, the approved content required for memory extraction may be sent to that provider. Check the selected provider, the material it can receive, and where its API credentials are stored before importing work conversations.

4. Decide how backup and deletion should work

The default data directory is ~/.one-memory. Removing the npm package only removes the Bootstrap CLI. memhub uninstall removes the managed service and runtime but preserves memory by default. The separate memhub delete-data command is the destructive path and requires the runtime to be stopped plus confirmation of the exact data directory. Start with a non-sensitive repository and define a backup policy before broadening the scope.

Install and verify the runtime

1. Inspect the release: Verify your OS and CPU against the v1.0.4 assets and its release manifest.
2. Install a matching Bootstrap CLI: Pin the version as shown above. npm installation alone does not download the full runtime or register a background service.
3. Review the plan: Run memhub install and read the source URL, install location, data path, and integrity checks.
4. Apply the plan: Run memhub install --yes only if the details match. A managed Linux background service requires a working systemd user session.
5. Initialize and diagnose: Use the commands below and resolve any health check before connecting a real session source.

memhub init
memhub status
memhub doctor

Apple Silicon Mac users can alternatively download the notarized v1.0.4 DMG. The desktop application and CLI share one runtime, data directory, and permission contract. Closing the Electron window is not necessarily the same as stopping the background runtime because the managed LaunchAgent can continue running.

MemHub imports authorized history, continues syncing new sessions, and lets connected tools retrieve relevant memory. Source: Tencent Hunyuan Hy Memory 2.0 page, checked October 20, 2026.

What the Codex and Claude Code connections change

MemHub's one-click connection is more than an account toggle. For Codex it adds an MCP entry to ~/.codex/config.toml, installs ~/.codex/skills/memhub-recall/SKILL.md, and adds UserPromptSubmit and Stop capture hooks. For Claude Code it adds the MCP server to ~/.claude.json, installs a recall Skill, and merges UserPromptSubmit, Stop, and SubagentStop hooks into ~/.claude/settings.json.

Recall and synchronization are separate controls. If the app connection is on but sync is off, the tool can query existing memory but new sessions will not enter MemHub. If sync is on but the connection is off, MemHub may continue reading approved session sources while the host no longer has a memory tool. Both controls must be on for two-way use.

The current integration does not guarantee that an already running Codex or Claude Code process hot-reloads its MCP, Skill, and hook files. Fully exit the client, start a new session, and verify that the MCP tools and managed Skill are available. If you have manually modified the same files, uninstall may revoke the credential and ask for a manual review rather than deleting your changes.

Three practical workflows

Carry repository rules between tools

In the first session, establish the package manager, test commands, files that must not be edited, and review criteria. In a second tool, ask it to retrieve and summarize the repository rules before proposing a plan. The first output should be a checklist, not a code change. Verify every remembered rule against the current README, package files, and CI configuration because memory can become stale.

Preserve decisions during a long migration

Across a multi-day framework upgrade, record the chosen version, rejected alternatives, failure causes, and remaining work. A new session can retrieve the relevant decision trail before suggesting a next step. Require it to compare remembered paths and branches with the current checkout. Memory is a navigation aid, not a replacement for Git history, issues, or an architecture decision record.

Hand off design and implementation across agents

You might use Claude Code for research and design, then Codex for implementation. Before coding, retrieve only the goal, constraints, and completion criteria. Exclude secrets, customer data, and unnecessary transcript text from the approved source scope. This keeps the handoff useful without turning all prior conversations into ambient context.

Privacy and troubleshooting checklist

• Debug logging: The default info-level design excludes conversation bodies and credentials. Debug mode can write full model prompts and responses to local logs. Use it briefly, switch back to info, and inspect or remove the old debug files separately.
• Sharing diagnostics: Remove API keys, real conversations, access tokens, and complete local paths before attaching logs to an issue.
• Connection not visible: Check runtime health, credential approval, the recall connection, the sync control, and then restart the host.
• New sessions are missing: Recall and sync are different. Confirm that Core scanning and capture hooks are enabled for that source.
• Installation stops: Check CLI/release version parity, actual assets for your OS and architecture, Node.js 24+, the Linux systemd user session, and proxy access to GitHub. Never bypass a failed checksum.

Bottom line

The most useful idea in Hy Memory 2.0 is not simply a larger memory model. It is the separation of memory from a single model and client, with independent controls for source ingestion and host recall. MemHub is still a preview, and its packages are moving quickly. For an evaluation today, pin a version such as v1.0.4 that has a real runtime manifest and artifacts, connect only a non-sensitive project, and test recall and synchronization as separate permissions.

Official sources: Hy Memory 2.0 announcement, Hy-Memory repository, MemHub releases, MemHub privacy document, Codex integration document, and Claude Code integration document. Checked October 20, 2026.

한국어 버전으로 읽기