Hermes AI Memory: Check What Survives a New Session

Hermes AI memory depends on which Hermes product and configuration you use. For Nous Research’s Hermes Agent, inspect its saved memory and project instructions, then start a fresh session to test what is available without the old conversation. This guide covers that agent, rather than assuming a model with a similar name stores your work.

fenbs can hold a separate project record for supported assistants and people who need to continue the work later. Keep that record explicit: the goal, decisions, completed checks, and next action should remain understandable even when the next assistant does not read your local Hermes files.

How do you run a useful first memory test?

1. Confirm the product is Nous Research’s Hermes Agent, and identify the active profile and configuration you are testing.
2. Choose a harmless project fact that is not obvious from the task, such as a fictional review label used only in this test.
3. Ask Hermes to save it. Inspect the saved entry or the tool result instead of treating an acknowledgment as proof.
4. Start a genuinely new session in the same profile. Ask for the label and where that information came from without repeating it.
5. Correct the fact, repeat the test, and confirm the fresh session uses the replacement. Remove the temporary test entry afterward.

Keep the test separate from valuable work. A made-up label lets you test persistence, meaning survival between sessions, without storing a password or customer information. Record the result alongside the profile and conditions, so a later configuration change does not make an old result look universal.

Which Hermes records have different jobs?

Hermes documents built-in persistent notes in MEMORY.md and a user profile in USER.md. Its file guide↗ (opens in a new tab) distinguishes those from project instructions and the agent’s personality file. The saved notes are loaded as a session-start snapshot; saving something during a conversation does not rewrite that original snapshot.

These are different checks. Recovering a phrase from an earlier conversation does not prove it lives in always-available memory. Following an instruction file does not prove a chat detail survived. Ask the assistant to identify the source, then inspect it yourself when the consequence matters.

Why can a project instruction seem to disappear?

Check the active directory and the instruction file selected for the session. Hermes has several compatible project context formats and a priority order, so a file you edited may not be the one loaded. Its context files documentation↗ (opens in a new tab) explains selection, directory scope, and overrides.

Use a harmless instruction to diagnose this, such as asking the assistant to identify the current project before discussing changes. Compare its answer with the loaded file. Avoid changing several competing instruction files at once: one controlled change makes it easier to see whether the problem is selection, stale text, or the assistant’s behavior.

Keep stable project guidance in the appropriate shared source. Temporary investigation details need a stopping-point note, rather than another permanent instruction. A portable AI memory record should explain the work to people and other tools without requiring them to reproduce the same local file selection.

How should you investigate a missing or wrong memory?

First inspect whether the entry exists, then whether it says what you intended. Next check whether you started the session in the expected profile. Finally ask the assistant to explain which evidence it used. This order separates a failed save from a failed retrieval and from a mistaken interpretation of a correct entry.

Hermes’s persistent memory documentation↗ (opens in a new tab) covers memory configuration, capacity management, editing, and session search. Use its current instructions for your installation rather than relying on remembered limits or commands. Session search can recover earlier discussion, while a concise saved entry serves a different purpose: keeping an important fact ready for later work.

Correct the smallest inaccurate statement you can identify. If the assistant inferred that a suggested workflow was approved, replace that assumption with the actual status and who must decide. Repeating the correction in the same conversation is not enough evidence; run the fresh-session test again.

What should survive when another assistant takes over?

Prepare a handoff that stands on its own. Say what the project is trying to accomplish, what changed, what was checked, and what remains uncertain. Name the current decision and the first next action. Include source references where they help someone confirm a claim, rather than asking the next assistant to trust an unsupported summary.

For example, a hypothetical website task might say the new form is implemented, keyboard navigation was checked locally, and email delivery is still untested. The next assistant should begin with delivery verification. “The form is finished” would hide that requirement and turn a handoff into an accidental approval.

fenbs’s AI handoff workflow keeps Where we left off with the tasks, so the next supported assistant has a defined starting point. Sharing that record with colleagues also needs the access and upkeep described in AI memory for teams.

How to do this with fenbs

1. Create a free account and add the project you want to continue across sessions.
2. Add its current task and a short AI context note containing verified facts. Record a person’s choices in Decisions and rules, with unanswered choices left open.
3. Save Where we left off with completed checks, unresolved questions, and the next action. A person can do this in the web app after a Hermes session.
4. For automatic assistant access, choose an app listed in Integrations, follow its connection guide, and ask it to read the project context before continuing.

fenbs is free for one person, no credit card. It does not import Hermes conversation history or replace Hermes’s own memory configuration. If you only need personal preferences within one local assistant, its native memory may be enough.

How do you keep the result dependable?

Repeat a small recall-and-correction test after changing profiles, storage, or your session workflow. Check whether the assistant can also admit a missing fact. An honest “I cannot find that” is a better starting point than a plausible invention, particularly when the next action changes a real project.

For the adjacent topics, read AI memory explained for the general idea and AI memory management for ongoing upkeep; the AI memory hub collects other workflows.