Deploy and Host Hermes Agent with Hindsight Memory on Railway

This template runs Hermes Agent by Nous Research with Hindsight by Vectorize as its long-term memory, both self-hosted in one Railway project. Hermes stores what you tell it in Hindsight, and Hindsight turns it into facts it can recall in later conversations. One OpenRouter key runs both.

Hermes supports Hindsight as a memory provider through a plugin that Vectorize maintains, but setting it up means installing the plugin, running Hindsight with Postgres and pgvector, and pointing one at the other. Here all of that is done on deploy. Hindsight has no public address at all; Hermes talks to it over Railway's private network with a generated key.

About Hosting Hermes Agent with Hindsight Memory

The deploy form asks for an OpenRouter API key. Set MODEL on the hermes-agent service too if you don't want Hermes's default model. When the deploy is done, use HERMES_API_URL and API_SERVER_KEY from the Variables tab of hermes-agent with any OpenAI-compatible client (the model name is hermes-agent), or add a Telegram bot with TELEGRAM_BOT_TOKEN and your user ID in TELEGRAM_ALLOWED_USERS.

What I tested before publishing: I told Hermes a made-up code word in one chat. Hindsight's logs showed it storing the fact in the hermes bank. In a second chat with no shared history, Hermes answered with the code word, and Hindsight's logs showed the recall that supplied it (0.3 seconds). Then I restarted Hermes and asked again in a new chat: same answer, again served by Hindsight.

At idle the three services used about 0.94 GB of RAM together (Hermes 392 MB, Hindsight 388 MB, Postgres 160 MB), roughly $9 a month.

Common Use Cases

A personal assistant on Telegram that remembers people, projects and decisions across months of conversations
An agent behind an OpenAI-compatible API, for Open WebUI or your own code, with memory that doesn't live in the chat history

Dependencies for Hermes Agent with Hindsight Memory Hosting

An OpenRouter API key. Hermes uses it for its chat model, and Hindsight uses it to extract facts (openai/gpt-4o-mini) and to embed them (openai/text-embedding-3-small).

Deployment Dependencies

Hermes Agent: https://github.com/NousResearch/hermes-agent
Hindsight: https://github.com/vectorize-io/hindsight
The Hindsight plugin for Hermes: https://github.com/vectorize-io/hindsight/tree/main/hindsight-integrations/hermes
Images: nousresearch/hermes-agent:latest, ghcr.io/vectorize-io/hindsight-api:0.10.1-slim, pgvector/pgvector:pg17

Implementation Details

On first start Hermes installs the Hindsight plugin into its volume and sets memory.provider to hindsight. The plugin reads the Hindsight address, key and bank name from the HINDSIGHT_* variables on hermes-agent, so there's no config file to edit.

Pick Hermes's model with care. Every message carries about 11,500 tokens of Hermes's own instructions and tools before your text. That's a fraction of a cent on openai/gpt-4o-mini, and much more on Hermes's default, Claude Opus. MODEL is applied on every start.

Don't change the embedding model once memories exist. Vectors from one model can't be searched with another.

Hermes can run shell commands in its container, and so can anyone holding API_SERVER_KEY. Keep the key private and TELEGRAM_ALLOWED_USERS to people you trust.

Why Deploy Hermes Agent with Hindsight Memory on Railway?

Memory only helps if the agent and its database are always on, and here they run next to each other on a private network. You pay a few dollars a month for the RAM, and the model bill depends on how much you use it.