Should you deploy Memoh?
Start with an expendable trial if your agent needs a persistent computer and conversations across services. Memoh's documented combination of managed workspaces, saved context and messaging access gives that trial a clear purpose. Our repository run does not establish that those features work together in production.
Memoh can host its built-in agent, Claude Code or Codex in managed workspaces. The README at our tested commit describes filesystem, desktop, browser, network and long-term memory capabilities, plus Telegram, Discord, Lark, WeChat and web access. Those are project-documented capabilities, not workloads we exercised.
What our sandbox actually recorded
Our run on September 18, 2026, at commit 8603825 covered dependency installation and repository tests, not a deployed agent workload. The attached sandbox record specifies the lab-node:22 image, 3 CPUs and 8 GB of memory. The harness stores each command's exit status, elapsed time and timeout status alongside output tails and parsed test counts.
Installation exited 0 after 14 seconds. The stored dependency measurements are 1,050 packages and 1,144 MB. The installation tail separately says Packages: +1049; that package-manager output is not the same field as the harness's stored dependency count, and we have not substituted it for that measurement.
The build step was skipped with the reason no script. Detection found no generic build script, but did find web:build. The skipped step therefore provides no build verdict for the web application or other components.
The test command exited 1 after 103 seconds without timing out. The attached stored parser summary records 1,959 passed, 9 failed and 0 skipped, with a reported total of 1,968. These figures do not establish exhaustive unique test coverage. We did not rerun the suite or reconcile a complete log.
The raw output separately names failures in desktop server-connection tests, Docker server-entrypoint tests and the web UI icon contract. The desktop excerpt shows an expected-call mismatch involving request options; the web excerpt reports an assertion that expected true and received false. The supplied tail does not establish one cause for all failures. Its test-file footer describes files, a different unit from individual test outcomes.
The MrKeyoor testing-methodology page explains how command outcomes and parsed counts reach our reviews. It is linked for provenance; its current procedure does not prove which methodology version governed this older run.
Deployment needs a trusted host
The deployment guide warns that the main service has privileged container access and should run only in trusted environments. That requirement matters before you choose a host or allow other people to create agents.
The deployment guide calls for Docker with Compose v2, Git and a configuration file created before startup. It also directs production operators to replace default passwords and secrets, configure HTTPS and a firewall, set resource limits and maintain backups. Model access requires your own credentials when using the documented API-key path.
The supplied documentation differs on service layout: the deployment guide lists five standard services, while the commit-pinned README describes separate Server and Channel services. Check the Compose configuration for the exact revision you deploy instead of treating either description as a verified service inventory for our run.
Our sandbox did not start the Compose deployment, create a bot, run Claude Code or Codex, deliver a chat message, exercise browser control, or verify persistence after restart. The audit field is null, and scan coverage is unknown. Neither installation success nor the repository tests constitute a security review.
Which alternatives fit a different job?
OpenHands Agent Canvas is a self-hosted control center for coding agents and automations. Its current README documents local, remote and cloud backends, including Docker options. Evaluate it when engineering conversations and scheduled or event-triggered coding work are your main requirement. We did not compare its reliability or resource use with Memoh.
Cua provides computers and desktop automation tools for agents you bring. Its current README separates Spaces, Driver, VM tooling, SDKs and evaluation tools. Evaluate those components when you need the computer or automation layer for an existing agent harness. This is a component-level alternative; the supplied extract does not establish it as a replacement for Memoh's messaging and memory platform, or establish a supported integration between them.
How to read the scores
These are editorial assessments of the supplied evidence, not performance or adoption measurements. Setup ease uses 1 for substantial manual assembly, 3 for a documented deployment with several operational requirements and 5 for a verified simple setup. Memoh scores 2 because its documented installation path still requires privileged hosting and production configuration, and our run did not verify deployment.
Documentation uses 1 for missing setup instructions, 3 for usable installation guidance and 5 for consistent operational guidance verified in deployment. Memoh scores 4 for its installation, configuration, production and troubleshooting instructions; the conflicting service descriptions prevent a top score.
Community uses 1 for evidence of a single contributor, 3 for several contributors and 5 for a release crediting at least ten contributors. The v0.20.0 release credits two named contributors and eight others, supporting 5 on this narrow participation rubric. It does not measure user adoption, support responsiveness or contributor retention.
Maturity uses 1 for no runnable evidence, 3 for a versioned project with a completed repository test command but unresolved failures and untested deployment paths, and 5 for clean repository checks plus verified operational workloads. Memoh scores 3. The v0.20.0 release and our tested commit are separate references; our run was not a test of the release tag.
What to verify before keeping durable work
Pin the revision you intend to use. On a trusted trial host, create an expendable bot, complete an agent task, restart the stack and check that its files and context remain available. Exercise the browser and each channel you need, then delete the bot and inspect storage reclamation. These are proposed acceptance checks, not results from our sandbox.