How AI Agents Remember

What How AI Agents Remember does

The repository analyzes the memory code of OpenClaw, Hermes Agent, nanobot, NullClaw and OpenFang, with a folder of articles for each project. Every analysis covers an architecture overview, the data model, how memories are stored and retrieved, lifecycle management such as consolidation and cleanup, and a guide for rebuilding the design in your own stack. Diagrams are drawn in Mermaid, so GitHub renders them without image files.

The Hermes Agent section has ten articles. They describe a cache-first design in which a frozen system prompt is built once per session, small MEMORY.md and USER.md files hold always-present facts, and full history lives in a state.db SQLite database searched with FTS5 through session_search. External memory providers such as Honcho, Mem0 and Hindsight plug in through a single MemoryManager, and skills act as procedural memory. The main README is written in Chinese, with an English version linked.

Key features

• Source-level memory analysis of five agent projects
• Ten articles on Hermes Agent covering built-in memory, provider plugins, session search and context compression
• Data model, storage, retrieval and lifecycle breakdowns for each project
• Replication guides for rebuilding each design
• Mermaid architecture diagrams
• A side-by-side comparison of the five systems

When to use it

• Learning how MEMORY.md, USER.md and session_search fit together in Hermes
• Comparing Hermes memory design with OpenClaw, nanobot, NullClaw and OpenFang
• Designing memory for your own agent using the replication guides

Who it is for: Developers and agent builders who want a code-level explanation of how agent memory works.

How it fits with Hermes Agent

A reference for Hermes Agent users and one of five projects it analyzes; it documents Hermes memory rather than adding anything to it.

Note: It is a documentation repository whose main text is in Chinese.

FAQ