ReMe/integrations/hermes_agent

ReMe memory provider for Hermes Agent

This plugin connects Hermes Agent to a running ReMe HTTP service. It recalls relevant memory before each model call and records each completed turn through ReMe's automatic memory job.

Prerequisites

• Python 3.11 or newer
• A working Hermes Agent installation
• ReMe installed with its core dependencies
• One ReMe workspace and endpoint for each Hermes profile that should remain isolated

ReMe search currently covers one whole workspace. Pointing multiple Hermes profiles at the same ReMe workspace therefore shares their recalled memory. Use a separate ReMe workspace and endpoint when profiles must be isolated.

Start ReMe

Start the HTTP service against a workspace dedicated to the active Hermes profile:

reme start \
  workspace_dir="$HOME/.reme-hermes-default" \
  service.backend=http \
  service.host=127.0.0.1 \
  service.port=2333

ReMe needs a working LLM configuration for automatic memory extraction. Its default search uses BM25, so embedding credentials are optional unless vector retrieval is enabled. Keep the service running while Hermes is active.

Install and configure

Hermes supports installing a plugin from a repository subdirectory:

hermes plugins install agentscope-ai/ReMe/integrations/hermes_agent
hermes memory setup

Select reme, accept http://127.0.0.1:2333 or enter the endpoint used above. Setup calls ReMe health_check and only replaces an existing provider config after the endpoint reports healthy. Then start a new Hermes session. Configuration is stored in $HERMES_HOME/reme.json, so every Hermes profile can point to its own ReMe workspace.

The file supports these optional settings:

{
  "endpoint": "http://127.0.0.1:2333",
  "request_timeout": 600.0,
  "recall_timeout": 5.0,
  "health_timeout": 2.0,
  "health_retry_seconds": 30.0,
  "shutdown_timeout": 30.0,
  "recall_limit": 5
}

Run hermes memory status to check that the provider is installed and configured. Starting a Hermes session performs a fresh endpoint health check.

Lifecycle and failure behavior

• prefetch calls ReMe search and returns only its recalled text. Hermes wraps that text in its protected memory-context block.
• sync_turn queues the completed user/assistant turn for a serial background writer, which calls ReMe auto_memory with a filename-safe ID derived from the Hermes profile and conversation.
• Cron, flush, and subagent contexts do not write conversational memory.
• A failed health check disables recall and recording until the retry cooldown expires. Retrieval and recording failures use independent cooldowns, so one action cannot disable the other while the ReMe service remains healthy.
• Recall uses its own short timeout so a slow ReMe search cannot stall the Hermes model call for the longer automatic-memory timeout.
• shutdown gives queued writes a bounded drain interval; ReMe remains an independently managed service. An idempotent process-exit hook uses the same drain path when a Hermes surface does not call provider shutdown directly.