Markdown Edit

Hermes Agent Memory Providers

Hermes Agent Memory Providers is a pluggable memory subsystem within Nous Research's autonomous Hermes Agent framework. While Hermes Agent includes a built-in file-based memory foundation (utilizing local MEMORY.md and USER.md markdown files), the memory provider interface enables users to attach specialized external memory engines for structured context capture, multi-hop recall, and cross-session knowledge synchronization.

graph TD
    Agent[Hermes Agent Core] <--> LocalBase["Default Local Foundation (MEMORY.md / USER.md)"]
    Agent <--> ProviderManager[Memory Provider Manager]

    subgraph "External Pluggable Memory Providers"
        ProviderManager <--> Honcho["Honcho (User modeling & plasticity)"]
        ProviderManager <--> Mem0["Mem0 (Cloud / local fact extraction)"]
        ProviderManager <--> OpenViking["OpenViking (Tiered L0/L1/L2 retrieval)"]
        ProviderManager <--> Holographic["Holographic (Local SQLite HRR algebra)"]
        ProviderManager <--> RetainDB["RetainDB (Hybrid BM25 + Vector)"]
        ProviderManager <--> ByteRover["ByteRover (Hierarchical markdown tree)"]
        ProviderManager <--> Hindsight["Hindsight (Knowledge graph reflection)"]
        ProviderManager <--> Supermemory["Supermemory (MCP context engine)"]
    end

How Hermes Memory Providers Work

External memory providers in Hermes Agent operate alongside (rather than replacing) the built-in foundation:

  1. Pre-Turn Context Injection: Before each agent turn, the active provider prefetches relevant memories based on the user's latest message and injects them into the system prompt context.
  2. Post-Turn Synchronization: After the agent responds or executes tools, the provider synchronizes conversation turns and extracts new facts asynchronously.
  3. Agent Memory Tools: Provides the agent with specialized tools to search, update, or clear stored memories.

Supported Providers Breakdown

Provider Storage Architecture Local / Open Source Unique Strengths
Honcho Relational / Dynamic User Graph Yes (Local Docker / SaaS) AI-native user modeling, dialectic reasoning, and theory-of-mind plasticity.
Mem0 Vector + Structured Fact Layer Yes (Local / Cloud) Automated fact extraction, deduplication, and user/agent/session scoping.
OpenViking Tiered Context Storage Yes (Self-hosted) Multi-tier L0/L1/L2 loading designed to reduce context token usage by 80–90%.
Holographic Local SQLite Embedded DB Yes (Local embedded) Uses Holographic Reduced Representations (HRR) vector-symbolic algebra with zero cloud dependencies.
RetainDB Hybrid Database Index Cloud / Self-hosted Multi-modal hybrid search combining dense vectors, BM25 keyword matching, and rerankers.
ByteRover Markdown Knowledge Tree Yes (Local / Cloud) Hierarchical knowledge tree stored in human-readable Markdown with tiered retrieval pipelines.
Hindsight Reflective Knowledge Graph Local / Cloud Extracts discrete entity facts and performs multi-hop reflective reasoning over conversational history.
Supermemory Knowledge Graph + MCP Connectors Yes (Local / Cloud) Universal context engine connecting bookmarks, documents, and external apps.

Configuration & Usage

Hermes Agent allows selecting and switching memory providers via CLI or YAML configuration:

Interactive Setup

hermes memory setup

Check Memory Status

hermes memory status

Manual Configuration (~/.hermes/config.yaml)

memory:
  provider: "honcho" # Options: honcho, mem0, openviking, holographic, retaindb, byterover, hindsight, supermemory
  sync_enabled: true
  prefetch_limit: 5