Understanding "Memory"
To get the most out of Hermes, you need to know that it has two different types of memory:
Short-Term (The Conversation)
Like a whiteboard. It remembers what you talked about 10 minutes ago, but when you start a /new session, the whiteboard is erased.
Long-Term (The Filing Cabinet)
Like a permanent record. These are facts, preferences, and rules that Hermes keeps forever, regardless of whether you start a new session.
Common "What If" Scenarios
"What if my agent forgets something we just talked about?"
If the conversation gets very long, the AI can sometimes "lose the thread" of the beginning. This is a limitation of how all AI works.
The Fix:
"Just a reminder: we are working on [Task X] and the goal is [Result Y]. Now, let's continue with..."
"How do I get my agent to NOT forget something?"
If there is a fact (like a client's preference or a specific business rule) that you want Hermes to know every single time you chat, you must move it from the "Whiteboard" to the "Filing Cabinet."
The Fix (Just tell it!):
"Please remember this for the future: [Insert Fact Here]. Save this to your permanent memory."
"What if the agent remembers something that is now WRONG?"
Since the "Filing Cabinet" is permanent, it might contain old info (like a client's old address) that needs to be updated.
The Fix:
"You have a note that says [Old Info], but that has changed. Please update your memory to [New Info]."
"What if it's acting weird or giving strange answers?"
Sometimes a conversation gets "polluted" with too many conflicting instructions.
The Fix:
Use the /new command to wipe the whiteboard clean and start fresh. Your permanent memory will still be there, but the confusing conversation will be gone.
"What if the agent can't reach my tool (Telegram, Nextcloud, calendar)?"
Connections can silently break: an expired app password, a revoked permission, or a network hiccup. The agent will usually tell you rather than guess — that's good behavior, not failure.
The Fix:
Ask it to run a test: "Check your Nextcloud connection now." If it fails, re-check the credentials in the Connection Hub recipes (step: "Secure the Credentials"). App passwords expire or get deleted — regenerating one fixes most connection errors.
"What if it's asking permission for everything?"
New setups often run in "ask-first" mode, where every tool use needs your approval. It feels slow, but it's a safety feature — the agent won't take real actions (sending emails, moving files) without a yes.
The Fix:
Approve the actions you trust, and for repetitive tasks tell it: "You have my standing approval to read files in this folder. Always ask before sending anything to a client." Grant trust by task type, not blindly.
"What if it edited the wrong file or sent the wrong thing?"
Act fast and stay calm — most actions are reversible, and your agent keeps a history of what it did.
The Fix (in order):
1. Say "/undo" to walk back the last change. 2. If it already went out, tell it exactly what happened: "You sent the wrong attachment to Jane. Draft a correction email now." 3. Then prevent a repeat: "Before you send or overwrite anything from now on, show me exactly what you're about to do first."
"What if I don't want my prompts stored by a provider?"
Every hosted AI processes your prompt on someone else's computer, and providers differ wildly in what they keep afterwards — some for minutes, some for up to 30 days, some indefinitely by default. "Zero data retention" exists, but it is usually a setting you must enable or a contract you must sign, not a default.
The Fix:
Before sending anything sensitive, ask the provider: "After you process my prompt, what do you keep, and for how long?" If the answer is a retention window you can't accept — or you can't get a straight answer — move that work to AI you host yourself. See the Sovereign AI track for the full walkthrough.
"What if compliance says our data can't leave the building?"
This is more common than you'd think: legal matters, medical records, financial files, or contracts that simply prohibit third-party processing. The good news: this constraint no longer means going without AI.
The Fix:
Run the AI where the data lives. A single modest server running an open model can handle private chat and document Q&A for a 10–50 person organization — prompts never leave your network. Keep public AI services for work that is genuinely non-sensitive, and keep this rule of thumb: "Client identifiers, legal, medical, financial → keep it sovereign."
Knowledge Check
Three quick questions to test yourself. Nothing is submitted; this is just for you.
1. Your agent forgot something from 30 minutes ago. What do you do?
2. How do you make the agent remember a fact permanently?
3. A permanent memory is now wrong. You should…