Most AI tools reset the moment you close the tab. Every new conversation starts from zero, so you re-explain your stack, your preferences, your context. It adds up to a lot of wasted time.
Hermes agents are different. They live on a persistent server. When you come back tomorrow, next week, or six months from now, the agent knows your name, your projects, your decisions, and what it was working on.
What "persistent" actually means here
Context windows, the amount of text a language model can process at once, are not persistence. A 200,000-token context window still resets to zero when you start a new session. The model does not remember the previous conversation unless you manually paste it back in.
Persistent memory is a separate system that runs alongside the model. Between sessions, a memory layer stores what the agent has learned about you, your projects, and its own task history. At the start of a new session, relevant memories are retrieved and loaded into context. The agent does not start from scratch. It picks up where it left off.
This is how Hermes Agent is designed to work, and it is the main reason people use it over a general-purpose chatbot.
Three memory types Hermes uses
The user model is a structured record of who you are: your technical preferences, your communication style, the projects you are running, your time zone, how you like results reported. This loads at the start of every session and shapes how the agent responds without you having to re-explain.
Skill Documents are procedural memory: records of how the agent solved a particular class of problem. When Hermes successfully scrapes a tricky website, processes a specific data format, or debugs a pattern of error, it can synthesize that experience into a reusable document. The next similar task takes less time and requires less guidance because the agent already knows a working approach.
Event memory is a timestamped log: tasks undertaken, decisions made, results achieved, failures encountered and how they were handled. This is what lets the agent tell you what it did last Tuesday, or explain why it made a particular call three weeks ago.
Why this requires a server that stays on
Memory stored in a local application disappears if you reinstall, switch machines, or the application crashes. For memory to be genuinely useful over months of agent use, it needs to live somewhere persistent: a server with its own storage, separate from your local environment.
This is why running an agent locally on your laptop is fundamentally limited for memory-intensive use. You get the capability during the session, but you lose the compounding effect that makes the agent genuinely useful over time.
Hivra keeps the memory on the agent's own cloud server, not on your laptop. If you switch to a new computer, your agent's memory is exactly where you left it. Nothing is tied to a local process. Backup coverage depends on the agent and provider and is not guaranteed, so keep an export of anything you cannot afford to lose.
Memory isolation across agent profiles
If you run multiple agent profiles, such as a research agent, a coding agent and a customer support agent, each profile has a completely separate memory store. They do not share context and cannot contaminate each other's working memory.
The one shared piece is optional account memory for Claude Code and Codex agents: notes you write once on the dashboard's shared memory page are copied into each new Claude Code or Codex agent you launch. Hermes agents do not receive it. After that, each agent's memory evolves on its own. Agents do not coordinate or share memory automatically; if you want one agent to use another's notes, you pass them across yourself.
Memory review and correction
Hermes keeps its memory in plain files (USER.md and MEMORY.md) rather than a hidden store. If the agent learned an incorrect assumption, such as a wrong belief about how a system works or a mistaken preference, you can ask it to correct or remove that memory directly. No need to wait for it to get overwritten through repeated correction in conversation.
You can also download those memory files from the agent's file explorer. Claude Code and Codex agents add an Export data link on the agent's page that downloads their chats and memory as one JSON file. If you ever want to move to a self-hosted setup or a different hosting provider, you are not locked in by accumulated state.
What's included
- Full conversation and task history retained across sessions
- User model lives in USER.md: goals, preferences, communication style, project context
- General knowledge lives in MEMORY.md: curated facts the agent reads at session start
- Skill Documents: agentskills.io open format, procedural memory that compounds over time
- Event log: timestamped record of every task and decision
- Optional Honcho integration: cross-session AI-native user modeling across tools
- Checkpoint and rollback support: the /rollback command reverts file changes
- Memory isolated per agent profile: multiple agents, no cross-contamination
- Shared account memory for new Claude Code and Codex agents, edited once in the dashboard
- Memory files you can download from the file explorer, so you are not locked in
Common questions
How long does persistent memory last on Hivra?
Your agent's memory stays on its server for as long as the agent exists. Backups are not guaranteed, so download an export of anything you cannot afford to lose.
Can I reset or wipe an agent's memory?
Yes. Ask the agent to remove a specific memory, or delete the agent to wipe everything. Selective removal is supported, so you do not have to wipe everything to correct one wrong memory.
Does memory work across different agent profiles?
Each agent has its own memory. Agents do not share memory automatically; the only shared piece is optional account memory, which is copied into each new Claude Code or Codex agent when you launch it.
What format is the memory export?
Hermes memory is plain Markdown files (USER.md and MEMORY.md), which you can download from the agent's file explorer. Claude Code and Codex agents also have an Export data link on the agent's page that downloads chats plus memory as a single JSON file with a versioned schema, so your data is yours to keep, read, or move.
Is the memory encrypted?
Connections to your agent use HTTPS. The memory files themselves live on the agent's server disk, and Hivra does not promise separate encryption for them, so do not keep secrets in agent memory.
Can the agent's memory get too large and slow things down?
Not usually. Hermes keeps its core memory files short and curated, and uses full-text search over past sessions to recall older work, so it does not load months of history into every session.