How a Local Memory Plugin Fills Hermes’ Missing Piece After It Tops OpenRouter

Hermes Agent reached 2.71 trillion tokens in a single day on OpenRouter, surpassing OpenClaw’s 2.45 trillion tokens and becoming the top consumer within three months of its open‑source release.

Missing Process‑Level Memory in Hermes

Hermes stores information in two markdown files under ~/.hermes/memories/: MEMORY.md – 2 200 character limit (≈800 tokens) for environment facts and learned knowledge. USER.md – 1 375 character limit (≈500 tokens) for user preferences and profile.

Both files are injected as a frozen snapshot at the start of each session, so the agent cannot relate facts to the sequence of actions that produced them.

Example: after three attempts to migrate a Pydantic model from dict() (v1) to model_dump() (v2), Hermes recorded only the fact “project uses Pydantic v2” in MEMORY.md. The crucial experience that dict() fails under v2 and must be replaced was not stored, illustrating Hermes’s “curation‑style” memory that keeps facts but discards process details.

MemOS Local Plugin 2.0 – Executing‑While‑Learning

MemTensor released MemOS Local Plugin 2.0, which adds a double‑layer feedback mechanism:

Step‑level feedback – After each execution step the environment’s objective outcome (e.g., command succeeded, file created) is recorded.

Task‑level feedback – After the task completes, the user’s subjective evaluation (explicit like/dislike or implicit by acceptance/rejection) is recorded.

Agent records each execution step, observation, reflection, and user feedback as auditable, attributable, reusable learning signals.

The two layers allow the memory system to distinguish between merely “runnable” solutions and those that are truly satisfactory.

Shared Core Between Hermes and OpenClaw

Both agents use the same host‑agnostic core called Reflect2Evolve . The architecture consists of:

Core algorithm and storage (Reflect2Evolve).

Stable agent‑contract defining the interface.

Adapters: TypeScript process for OpenClaw, Python MemoryProvider + JSON‑RPC for Hermes.

Because the core and storage are shared, experiences learned by one agent become instantly available to the other, demonstrating “portable memory assets.”

Installation

For an existing Hermes installation, add MemOS with a single command:

curl -fsSL https://raw.githubusercontent.com/MemTensor/MemOS/main/apps/memos-local-plugin/install.sh | bash

The script detects the installed agent, installs the appropriate adapter, and leaves the original MEMORY.md and USER.md untouched while providing a separate viewer for the enriched memory.

Cold‑start latency : Little difference in the first week; benefits become noticeable after two weeks as traces and skills accumulate.

Viewer overload : New users should initially focus on memory and skill tags; trace and policy sections are useful for debugging specific behaviors.

Conclusion

Hermes’s 3 575‑character memory can store static facts about the user and project but cannot retain the step‑by‑step process that produced those facts. MemOS supplies the missing L1 (step‑level), L2 (pattern abstraction), L3 (global context), and Skill layers, turning the agent into an AI colleague that learns within the local environment.

General‑purpose models (GPT‑4, Claude Sonnet 4.5, Gemini 2.5, etc.) already provide strong reasoning; the next decisive factor for agents is the ability to acquire and retain process‑level knowledge locally.

Open‑source repositories:

GitHub: https://memos-docs.openmem.net/cn/openclaw/local_plugin

Technical docs: https://memos-docs.openmem.net/openclaw/hermes_local_plugin