> ## Documentation Index
> Fetch the complete documentation index at: https://docs.wazoo.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Hermes Agent Memory

> Worlds vs Hermes Agent Memory by Nous Research: multi-tiered LLM memory vs developer-native RDF graph context.

## At a glance

|              | Worlds                                           | [Hermes Agent](https://github.com/NousResearch/Hermes-Agent) (Nous Research)             |
| :----------- | :----------------------------------------------- | :--------------------------------------------------------------------------------------- |
| Data model   | RDF triples in an append-only fact ledger        | Four-tier memory (Markdown files, SQLite FTS5, Procedural Skills)                        |
| Memory tiers | Curated RDF knowledge graph                      | Prompt memory (`MEMORY.md`), Session archive (`state.db`), Procedural skills (`skills/`) |
| Retrieval    | Hybrid: vector, full-text, RDF filters           | FTS5 full-text search (`session_search`) & system prompt injection                       |
| Provenance   | Fact-level, chronological, verifiable via SPARQL | Session/file-level tracking                                                              |
| Deployment   | Edge-ready adapters (LibSQL, Postgres) or Cloud  | Local CLI framework / agent runtime                                                      |

## The philosophical difference

The [Hermes Agent](https://github.com/NousResearch/Hermes-Agent) architecture
(developed by Nous Research) organizes agent memory into a pragmatic four-layer
stack:

1. Prompt Memory: a static snapshot of `MEMORY.md` and `USER.md` injected into
   system prompts.
2. Session Archive: historical interaction logs indexed via SQLite FTS5
   (`state.db`) for episodic keyword recall.
3. Procedural Memory: extracted task procedures saved as executable markdown
   skills.
4. Pluggable Memory Providers: hooks to external memory stores (e.g. Mem0,
   Supermemory).

Hermes optimizes for agent self-curation and episodic session history within a
local agent CLI runtime.

Worlds approaches context from an open-standard, graph-first stance. Rather than
relying on flat markdown files or SQLite FTS search alone, Worlds enforces RDF
triples where every fact has explicit relationships, temporal validity, and
deterministic SPARQL queryability.

Where Hermes is an agent runtime managing its own prompt files and SQLite
session database, Worlds is a dedicated context engine that Hermes agents (or
any other agent runtime) can query as a verifiable memory provider.

## When Worlds fits

* You need deterministic [graph querying](/worlds/query) (SPARQL) to traverse
  relationships across complex domain facts.
* You require fact-level chronological provenance and auditability across all
  asserted data.
* You want an open, standards-based RDF memory layer shared across multiple
  agent runtimes (Claude Code, Letta, Hermes).
* Your application demands strict compliance, safety, and deterministic
  filtering where keyword or similarity search alone is insufficient.

## When Worlds does not fit

* You want an all-in-one local agent CLI framework that manages prompt context
  files (`MEMORY.md`) and session history out of the box.
* You prefer simple FTS5 text search over episodic agent conversations without
  establishing a formal knowledge graph vocabulary.

## Coexistence

Hermes agents support pluggable memory provider plugins. Hermes can use its
internal `MEMORY.md` for local prompt conventions while connecting Worlds as an
external memory provider tool to query authoritative RDF facts via SPARQL during
complex multi-step reasoning.