Memory is what separates a clever demo from an agent you actually rely on. Without it, every session starts from zero. The agent re-asks what you told it yesterday and re-derives what it already figured out. A persistent-memory skill records facts, decisions, and learned patterns to durable storage and pulls them back when relevant, so the agent compounds instead of resetting.
This is the single clearest place OpenClaw and Hermes take different philosophies, which makes it worth understanding before you commit to one.
How does memory work on OpenClaw?
OpenClaw treats memory as a category of skills you choose from, not a single built-in. The popular options on ClawHub include Ontology (structured, graph-style memory), braindb ("persistent, semantic memory for AI agents"), and agent-memory-ultimate (a production-ready system with daily logs and sleep-style consolidation). You install whichever matches your needs:
openclaw skills install ontologyThe upside is control. You decide the storage backend, recall strategy, and how aggressively old memories are pruned. The trade-off is that you own those decisions.
How does memory work on Hermes?
Hermes builds memory into the framework. The agent maintains agent-curated memory with periodic nudges, does full-text (FTS5) search over past sessions with LLM summarization for cross-session recall, and, as of recent versions, runs an autonomous Curator that grades, consolidates, and prunes its skill and memory library on a regular cycle. You don't install anything; it's how Hermes works out of the box.
If you outgrow the defaults, you can layer community memory skills on top, stronger cross-session user modeling, retain/recall/reflect workflows, or portable shared-memory artifacts, but most people never need to.
OpenClaw vs Hermes: which memory model should you pick?
- OpenClaw: modular; pick Ontology / braindb / agent-memory-ultimate and tune it yourself.
- Hermes: native; built-in learning loop, session search, and a self-pruning Curator.
Choose OpenClaw if you want to design the memory system; choose Hermes if you want one that already works and improves itself.
Can you manage agent memory from your phone?
Memory lives on the host running the agent, so managing it means reaching that host. Onepilot gives you SSH access from an iPhone to inspect the memory store, run a manual consolidation, or back it up, whether you're on OpenClaw or Hermes. Pair it with the Obsidian skill to make that memory human-readable.