Supermemory for Hermes Agent
The native Hermes provider adds external recall and capture alongside built-in memory. Here is what to configure and what to verify.
Supermemory is available as a native memory provider for Hermes Agent. It connects an agent's ongoing work to external memory search and profiles while leaving Hermes's built-in memory available.
The useful question is whether a returning session receives the right earlier context. A project decision should help the next relevant task without being mistaken for a permanent instruction or leaking into an unrelated workspace.
Hermes already has memory
Hermes includes MEMORY.md and USER.md. An external provider is an additional option when you need searchable context and explicit scopes beyond the local file workflow.
Use the Hermes setup guide for the current configuration and the upstream provider README for version-specific behavior.
Profiles and retrieval have different jobs
A profile can supply reusable background such as a preference or current project. Search can retrieve evidence relevant to a particular question. Use source references and updates to keep that context current.
For example, a user might say they are migrating a service to PostgreSQL. On a later related task, the agent can retrieve that decision and its source. It should still inspect the current repository before assuming the migration is complete.
Capture runs through the provider
When enabled, the current provider captures completed turns and retries pending writes at later lifecycle events. Pending writes can outlive the session that produced them, so check their status after an interruption.
Its upstream documentation explicitly describes at-least-once retry behavior. A lost response can lead to a repeated append. Test a network failure and inspect stored records for duplicate turns.
Eligible additions to native memory can also be mirrored. Verify existing files, later edits and deletions separately when deciding which store should supply a fact.
Work and personal context need explicit settings
The default container is hermes. To separate Hermes profiles, use the documented hermes-{identity} template. Optional custom containers provide further namespaces for explicit tool calls.
Automatic capture uses the primary container. Configuring work and personal containers does not make the model reliably classify every conversation into the correct destination. Inspect actual writes and use separate profiles or deployments when the required boundary calls for them.
Start with one fact across two sessions
Install the SDK and select the provider:
pip install supermemory
hermes memory setupUse a Supermemory console key, save a fictional project decision, and retrieve it from a new session. Check the source, update the decision, and verify the correction. Repeat with another scope to confirm it does not receive the same private context.
The coding plugin is free; service usage follows current pricing. Model-provider charges remain separate.
Cross-client context requires matching access
A shared service can make selected records available to other configured clients. It does not automatically move every Hermes conversation into ChatGPT, Claude Code or another product. The clients need compatible access, the intended scope, and an actual retrieval step.
For personal or team clients, the MCP context-sharing guide shows a controlled save-and-retrieve test. Check the account and scope in each client before relying on shared context.
Keep lifecycle controls visible
A changed fact is not necessarily a deleted source. Use the memory lifecycle guide to define correction, forgetting and deletion. Check conversation history and local files as well as external search when a removed fact returns.
Configure Supermemory for Hermes when that workflow fits your use case. Begin with inspectable evidence and scope checks, then expand capture after the behavior is clear.
Frequently asked questions
Does the provider automatically route every conversation into work or personal memory?
No. Automatic operations use the primary container. Allowed custom containers can be selected on explicit tool calls.
Does session capture guarantee no lost or repeated content?
No. The current provider documents at-least-once retries and has a bounded pending buffer. Verify the behavior under failure.
Can other clients use the same stored context?
Only when they have compatible access and use the intended scope and retrieval path. Installing a second client does not automatically transfer its conversations.