Zep Memory Provider

Persistent, cross-session memory powered by Zep — a context engineering platform that builds temporal knowledge graphs from conversations and business data.

What it does

• Persists every conversation turn to Zep's knowledge graph via sync_turn
• Prefetches a context block (user summary + relevant facts) before each agent response
• Exposes zep_search and zep_add tools so the agent can query and write to the graph
• Mirrors built-in MEMORY.md writes to Zep for unified knowledge
• Warms the Zep user cache between turns for low-latency retrieval

Setup

1. Install the Zep Python SDK:

   pip install zep-cloud

2. Run the Hermes memory setup wizard:

   hermes memory setup

   You'll be prompted for your Zep API key (get one at https://app.getzep.com).

3. Optionally configure user_id, first_name, and last_name in $HERMES_HOME/zep.json.

Config

Key	Description	Default
api_key	Zep API key (stored in .env)	—
user_id	Zep user ID for this profile	hermes-user
first_name	First name on the Zep user node	Hermes
last_name	Last name on the Zep user node	User

Tools

Tool	Description
zep_search	Search the knowledge graph for facts or entities
zep_add	Add text or JSON data directly to the user's knowledge graph

CLI commands

When Zep is the active provider:

hermes zep status   # Check connection and user info
hermes zep config   # Print current config
hermes zep search "query"  # Search the knowledge graph

How it works

Zep automatically extracts entities, relationships, and facts from conversations, building a temporal knowledge graph per user. The prefetch hook calls thread.get_user_context() each turn, which returns a context block containing a user summary and the most relevant facts — ready to inject into the system prompt. Facts include temporal validity ranges so the agent can reason about what's current vs. outdated.

For more details, see the Zep documentation.