Tencent publishes MemHub CLI for shared AI memory

Tencent has published MemHub CLI on npm, adding another installation route for its AI memory software. The official October 18 announcement pairs memhub-cli 1.0.5 with MemHub v1.0.4. This release brings a practical question for developers who want to carry useful context between AI tools. What gets installed, and where does the information go?

The npm registry records version 1.0.4 appearing on October 18, followed by 1.0.5 less than 23 minutes later. The broader project was already available. MemHub v1.0.0 was published on October 16. The fresh development is the public npm distribution and the accompanying runtime update.

The CLI is only the beginning

The product guide describes a local store built from AI conversations and coding records that the user authorizes. Supported AI tools receive separate permission to read that memory. Desktop and terminal interfaces share the same runtime and data directory. Tencent still describes MemHub as a preview, so those capabilities should be evaluated against the specific client and platform a team uses.

Installing the npm package supplies a bootstrap tool. According to the installation guide, the runtime download and service setup happen through a separate memhub install step, with a plan shown before confirmation. The installer checks the release manifest, operating system, architecture, protocol compatibility and file hashes. The documented sequence then uses memhub init and memhub status.

The package requires Node.js 24 or newer. The registry also marks version 1.0.4 as deprecated because its default GitHub runtime download encounters a redirect allowlist problem, and directs users to 1.0.5. Anyone who grabbed the first npm version should check which version they have before troubleshooting the rest of the setup.

The subsequent source change adds the relevant GitHub download hosts and separates the CLI version from its runtime target. ByteForward also inspected the published 1.0.5 package without executing it. That package explicitly targets runtime 1.0.4, explaining why the two version numbers differ.

Choose from the artifacts that shipped

The v1.0.4 release and its download manifest list a desktop app and runtime for Apple silicon Macs, plus runtimes for Linux x64 and arm64. No Intel Mac or Windows artifact appears in that set. The desktop manifest specifies macOS 13 as its minimum. Check that actual download list before treating a platform mentioned elsewhere in the documentation as available.

The lifecycle documentation also separates removing the npm command from removing the runtime. Uninstalling the npm package leaves the managed runtime and memory in place. The runtime uninstall command normally preserves memory too. Deleting the stored data is a separate operation requiring confirmation. Teams evaluating the tool should include cleanup in their trial plan.

Local memory still needs a data policy

The privacy guide says authorized content needed for a task is sent to the third party model service the user chooses. Local storage therefore does not settle where model processing happens. Review the selected provider and the information it will receive before connecting work conversations. Source access, model processing and access by AI tools have separate permission boundaries.

Debugging deserves particular care. The same privacy guide says normal diagnostics omit conversation bodies, while explicitly enabled debug mode can write full model prompts and responses into local logs. Switching back to normal logging stops further recording but does not remove those existing files. Inspect and redact a diagnostic bundle before sharing it.

That review extends beyond memory tools. ByteForward’s AgentPlay coverage explains how coding session records can remain revealing even when an exported view omits the full conversation. For a MemHub evaluation, start with material you are allowed to process and inspect the retrieved memory before relying on it in another tool.

This report is based on public package records, release artifacts, source changes and documentation checked on October 18. ByteForward did not install or run MemHub, test its integrations or measure memory quality.

Illustrative programming workspace from 2009. Working lat(t)e by Kuba Bożanowski via Wikimedia Commons under Creative Commons Attribution 2.0. Cropped and resized for display.