How Hermes Agent’s Memory System Fixes the Layered Misconception in OpenClaw

Hermes Agent introduces a four‑layer memory system that separates high‑value facts, historical conversations, procedural knowledge, and deep user models, avoiding the “big memory bag” design of many agents. The hot memory files MEMORY.md (2,200 characters) and USER.md (1,375 characters) store only information that must be present in every round, such as user preferences, environment facts, and stable conventions. These files are injected as a frozen snapshot into the system prompt at session start, protecting prompt‑caching efficiency.

Hot Memory vs. Archive

Hot memory ( MEMORY.md and USER.md) holds short, high‑frequency items that remain stable across rounds.

Historical sessions are stored in a SQLite database ( ~/.hermes/state.db) with FTS5 full‑text search. The session_search tool retrieves relevant past dialogs, aggregates by session, truncates, and summarizes before feeding the result to the main model.

Procedural knowledge lives in ~/.hermes/skills/ as “skills” (procedural memory). Examples include PR review checklists, deployment‑failure troubleshooting steps, and data‑export pipelines.

Optional external providers (Honcho) enable deep user modeling and cross‑device continuity without polluting the stable prompt prefix.

When a session is about to be compressed, Hermes performs a memory flush : it prompts the model to extract durable facts (user preferences, recurring fixes, high‑value information) and writes them back to hot memory before the history is shortened. After compression, the system prompt is rebuilt so the newly flushed facts become part of the next round’s stable context.

Prompt Ordering and Safety

Default Agent Identity
Tool usage guidance
Optional Honcho block
Optional system messages
Frozen MEMORY.md snapshot
Frozen USER.md snapshot
Skills index
Context files (AGENTS.md, SOUL.md, .cursorrules)
Date, time, platform
Conversation history
Current user message

This ordering keeps stable prefixes at the front, ensuring high cache‑hit rates. New memory entries are written to disk but do not immediately alter the current prompt, preserving stability. The memory tool only supports add, replace, and remove actions using substring matching, avoiding the need for internal IDs.

Security checks reject duplicate entries and dangerous content (prompt injection, credential leaks, hidden Unicode) because anything written to memory may later appear in the system prompt and affect model behavior.

Comparison with OpenClaw

OpenClaw emphasizes a large, workspace‑driven memory plane and extensive gateway controls, while Hermes focuses on a cache‑aware runtime: tiny hot memory, a searchable archive, procedural skills, and optional external providers. Both aim to prevent agents from relying on ever‑growing chat histories, but Hermes explicitly asks “which layer should this information belong to?” before committing it.

Practical guidance for building a custom agent memory system includes:

Identify which facts merit hot memory (user preferences, stable facts).

Store full conversation logs in an archive searchable by keywords and session IDs.

Before compressing long sessions, flush durable state to hot memory.

Encode repeatable procedures as skills, version them, and allow updates or removal.

Consider external user models (Honcho) only after hot memory, archive, and skills are solid.

Instrument the system to log what enters the prompt, what is retrieved from the archive, and what skills are invoked.

Overall, Hermes demonstrates that a well‑structured, layered memory plane reduces prompt cost, improves cache stability, and makes long‑running agents more reliable.