A personal AI agent runs continuously, retains memory across sessions, and operates inside your apps. You interact with it via chat, email, or voice. It continues working after you close the window.
The category began small. OpenClaw started in November 2025 as Clawdbot, a weekend project by Peter Steinberger. Twelve months later, Google, OpenAI, Meta and Perplexity all released their own versions.
The products split into two camps. Self-hosted agents run on your machine or server and use any model you choose. Managed agents run on a vendor’s cloud VM and come bundled with a subscription.
How we ranked them
We scored each agent on three axes that matter once an agent holds your inbox and credentials.
- Memory: what persists, where it is stored, and whether you can read, edit or delete it.
- Security: default permissions, the isolation boundary, approval gates, and the public incident record.
- Cost: the subscription price, or token spend plus hosting for self-hosted options.
Facts come from vendor documentation, launch posts and independent security research. Prices are US list prices as of November 21, 2026. Ranking reflects overall fit for an individual user, not raw capability.
Comparison table
| Rank | Agent | Maker | Runs on | Memory | Security default | Starting cost | Best for |
|---|---|---|---|---|---|---|---|
| 1 | Hermes Agent | Nous Research | Your machine or server; Docker, Modal, Daytona backends | MEMORY.md, USER.md, SQLite FTS5 session search, self-written skills | Manual approvals, denied on timeout; hardline denylist | Free (MIT) + model tokens | Developers who want an agent that learns |
| 2 | OpenClaw | OpenClaw Foundation | Your machine or server; opt-in sandbox | Editable Markdown workspace files; SQLite sessions since 2.0 | Full host access unless sandbox is enabled | Free (MIT) + tokens; managed hosts ~$12–$49/mo | Widest chat-channel reach |
| 3 | Gemini Spark | Dedicated Google Cloud VMs | Google account context via connected apps | Asks before spending money or sending email | $19.99/mo AI Pro (US) | Google Workspace users | |
| 4 | OpenAI dots | OpenAI | OpenAI cloud computer; optional connected computer | Chats, ChatGPT memory and the dot’s own notes | Proactive research is read-only; passwords hidden from model | Included in eligible ChatGPT Pro and Business Premium plans | ChatGPT and Codex power users |
| 5 | Meta Muse | Meta | Muse Secure VM in Meta’s cloud | Learns preferences; forgets on request | Sentinel agent gates internet access; approvals for email and purchases | Free; $20 and $100 tiers | Everyday errands on a free tier |
| 6 | Perplexity Personal Computer | Perplexity | Your Mac plus Perplexity cloud | Not detailed publicly | On-device authorization, 2FA, audit trail | Pro or Max subscription | Mac users who need local files |
| 7 | NanoClaw, IronClaw, ZeroClaw | Open-source projects | Containers, WebAssembly or a small Rust binary | Varies by project | Isolation on by default | Free + tokens | Security-first self-hosters |
1. Hermes Agent (Nous Research)
What it is: Hermes Agent is an open-source agent from Nous Research.
It runs from a terminal UI, a native desktop app, a local web dashboard and messaging gateways. Parallel counts about 25 supported platforms as of October 2026, including Telegram, Slack, WhatsApp, Signal, Teams and email. Native iMessage arrived in June 2026.
Memory: Memory is the main reason to pick Hermes. It keeps persistent MEMORY.md and USER.md files and curates them with periodic nudges. Session history sits in SQLite with FTS5 full-text search and LLM summarization. Pluggable providers include Honcho user modeling plus 8+ community backends.
Hermes also writes its own skills. It turns completed complex tasks into reusable skills and refines them over time. Skills follow the agentskills.io standard. The /learn command captures a workflow as a slash command without hand-writing SKILL.md.
Security: The docs describe a 7-layer defense-in-depth model. Layers include user allowlists, DM pairing and human approval for destructive commands. Approval mode can be manual, smart (an auxiliary LLM rates risk) or off.
The defaults fail closed. An unanswered approval prompt is denied after 300 seconds. Your own deny rules survive even YOLO mode.
Nous is explicit about the limits. Its docs say the denylist does not sandbox a hostile or compromised agent. The project’s SECURITY.md names OS-level isolation as the one load-bearing boundary. For broad tool access, use a container backend such as Docker, Modal, Daytona or Vercel Sandbox.
Cost: The agent is free. You pay for model tokens from any provider: OpenRouter, OpenAI, Anthropic, or local runtimes like Ollama and vLLM. Nous Portal is optional, with Free, $20, $100 and $200 monthly tiers. Local models cut token spend to zero; see the one-click local model setup.
Comments: Hermes is the best open-source pick for most developers. It pairs the deepest memory stack with fail-closed defaults. Setup is still terminal-first, and the local backend offers no protection from a compromised agent.
2. OpenClaw (OpenClaw Foundation)
What it is: OpenClaw defined the category. It started in November 2025 as Clawdbot and went viral in January 2026.
OpenClaw 2.0 shipped as v2026.8.1 on August 31, 2026, built by 933 contributors. The latest release we covered is 2026.9.5, with atomic updates and plugin hot reload.
A central Gateway process routes messages, runs cron jobs and hosts a Control UI. OpenClaw reached 26 channels as of v2026.6, including WhatsApp, Telegram, iMessage, Teams, Matrix and Signal. Supported models include OpenAI, Anthropic, Google, DeepSeek, xAI, Kimi, Ollama and OpenRouter.
Memory: OpenClaw stores behavior in human-editable Markdown files. SOUL.md holds the persona, TOOLS.md the tool rules and USER.md what the agent knows about you. Version 2.0 moved session and transcript storage into SQLite for better persistence.
This design is transparent. You can open, diff and back up every memory file with ordinary tools.
Security: This is OpenClaw’s weak point. The default mode gives full host access, which WIRED called incredibly risky. An opt-in sandbox runs sessions in per-session Docker containers. It defaults to no network egress, a read-only root and all capabilities dropped.
The incident record is long. CVE-2026-25253 allowed one-click remote code execution via a WebSocket origin bypass. Visiting a malicious page was enough to steal auth tokens. It was patched quickly.
The skills registry was the bigger problem. Koi Security audited 2,857 ClawHub skills and found 341 malicious entries. 335 traced back to one campaign, now tracked as ClawHavoc. Antiy CERT later counted 1,184 malicious skills as the registry grew.
OpenClaw responded. It now scans every uploaded skill with VirusTotal and re-scans daily. Version 2.0 added protected credential requests and a team-scoped Secret Store. Treat ClawHub as untrusted code anyway.
Cost: OpenClaw is free under MIT. A survey of managed hosts found plans from about $12 to roughly $49 per month. Watch the heartbeat check. With expensive models, it can drive significant token spend.
Comments: Pick OpenClaw for channel breadth and the largest ecosystem. Turn on the sandbox, skip unvetted skills, and keep the gateway on loopback. For a deeper head-to-head, see our OpenClaw vs Hermes Agent comparison.
3. Gemini Spark (Google)
What it is: Google unveiled Spark at I/O on May 19, 2026. It runs on Gemini 3.5 Flash with Google’s Antigravity harness. Each user gets dedicated VMs in Google Cloud, so tasks continue with your devices off.
Spark works natively across Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube and Google Maps. MCP connections extend it to third-party apps. You can message it by text and email, and a macOS app shipped in June.
Memory: Spark’s memory is your Google account. It pulls context from the apps you connect through Personal Intelligence. Google has not documented a separate, editable memory store like the open-source options.
Source Read original →

