Please use the form below to provide your feedback. Because your feedback is valuable to us, the information you submit in this form is recorded in our issue tracking system (JIRA), which is publicly available. You can track the status of your feedback using the ticket number displayed in the dialog once you submit the form.

Agent Memory Plug-in for Hermes Agent

  • concept
The Agent Memory plug-in adds persistent, semantic memory to Hermes Agent, recalling and capturing context automatically.

Hermes Agent is an open source autonomous AI agent built by Nous Research. Hermes' pluggable memory system lets you add Couchbase Agent Memory as a memory provider through the Couchbase Agent Memory plug-in. Once connected, the plug-in recalls and injects relevant memories into every turn and captures each exchange in the background, without slowing down Hermes. Captured exchanges go to your own Agent Memory server, not a third-party service.

To install the Agent Memory Plug-in for Hermes, see Install the Agent Memory Plug-in for Hermes.

Why Use Agent Memory with Hermes?

Hermes has a built-in memory tool that saves notes to MEMORY.md and USER.md. It only captures what the model decides is worth saving, and those files are limited in size. For example, if you mention that your team calls your staging environment preprod, the model may not recognize it as worth saving. When the conversation ends, this information is gone.

The Agent Memory plug-in solves this with a persistent, semantic memory layer that captures every turn. In a later conversation about something else, you can ask Hermes about the staging environment, and it still calls it preprod. You never have to repeat it or paste the context back in.

The Memory Loop

The plug-in works at 3 points in every turn:

1. Before the Turn

When you send a message, Hermes uses the plug-in to recall memories relevant to the upcoming turn in the background and adds them to the context for that turn. With these memories, Hermes has the context it needs without the model calling any tool. Should the prefetch take too long, the conversation continues with the model able to call the search tool during the turn.

2. During the Turn

When the model responds, it can call the plug-in’s tools to search, add, update, delete, or list memories directly. For more information about these tools, see Tools Exposed to the Model.

Both automatic recall and tools are on for every turn.

3. After the Turn

After the model responds, the plug-in captures the exchange by sending it to Agent Memory in the background, without blocking the next turn. From this exchange, Agent Memory extracts facts, generates embeddings, and summarizes it. An extracted fact might be user resides in London or user prefers Python. The exchange can be recalled on the next turn. The extracted facts can be recalled once processing finishes, typically within seconds.

Hermes Sessions

Hermes sessions map directly onto Agent Memory sessions:

  • Ending a Hermes session, either by running /reset or exiting the command-line interface (CLI), ends the corresponding Agent Memory session.

  • /resume and /branch continue the correct Agent Memory session, and resuming a conversation whose session was already ended continues it under a new, linked session.

  • By default, a search spans all of a user’s sessions.

Annotations

The plug-in automatically annotates every memory it writes with context from Hermes:

Annotation Values Meaning

source

turn, agent

turn marks an automatically captured conversation turn, and agent marks a fact the model stored explicitly through a tool call.

platform

cli, or a gateway name such as telegram, discord, whatsapp, signal, api_server

The Hermes surface the message arrived on.

agent

profile name

The active Hermes profile.

user_name

free-form

The sender’s display name, as reported by the gateway.

channel_id

free-form

The gateway’s chat or channel identifier.

channel_type

dm, group, forum, channel, thread

The kind of chat, as reported by the gateway.

An annotation appears only when Hermes provides it, so a plain CLI chat typically carries only platform, agent, and source.

For more information, see Narrow Recall for Every Search.

Hermes' Built-in Memory

Hermes uses a built-in memory that stays active by default, and installing the Agent Memory plug-in does not automatically turn it off.

Running both systems together, the model can keep saving facts to the built-in memory it already knows, so those facts never reach Agent Memory. To have a single source of truth, you can turn off Hermes' built-in memory.

For more information, see Install and Activate.

Multi-User Support with Hermes Profiles

Hermes identifies the active user with a single user_id, but changing that value alone is not enough to isolate 2 people on the same machine. Hermes' built-in memory ties to the active Hermes profile, not to the plug-in’s user_id. If you only switch the user_id, one person’s built-in notes can leak into another person’s conversations. Turning off built-in memory does not fix this either. Hermes also keeps session transcripts, /resume, and session search scoped per profile, and those cannot be disabled.

To run Hermes and maintain separation between people on a single machine, add 1 Hermes profile per person. A new profile is a fully separate home directory. It starts with no model configuration and no plug-in installed, so you need to set up both again for each profile.

For more information, see Running Multiple Users.

Tools Exposed to the Model

The Agent Memory plug-in provides the following tools to the model:

Tool Arguments Purpose

agentmemory_search

query, top_k (default 10), since (optional recency window, for example 24h, 7d, 2w, 3mo), plus the optional recall facets described in Narrow Recall for a Single Search

Semantically search the user’s memory, optionally narrowed by recency or facet.

agentmemory_add

content

Store a fact verbatim.

agentmemory_update

block_id, content

Replace a memory’s content. If the memory’s original session has already ended, this moves the content to a new block and returns its block_id. Use that new block_id for any further edits.

agentmemory_delete

block_id

Delete a memory.

agentmemory_list

limit (default 20, maximum 200)

List stored memories, most recent first.

See Also