Installing MemOS on Hermes: How a Memory Plugin Enables Self‑Generated Skills

MemOS Overview

MemOS is an open‑source “memory operating system” that provides long‑term memory for AI agents such as Hermes and OpenClaw. It stores memory in structured layers and offers a visual memory panel for inspection and management.

Memory Design Comparison

Claude Code : stores project experience and work rules in markdown files; loaded at session start.

OpenAI ChatGPT : keeps user profile and cross‑chat personalization in internal platform state; context is synthesized dynamically.

OpenClaw : uses MEMORY.md, daily notes, and an index as an auditable knowledge base; injected at startup and retrieved on demand.

Hermes Agent : limits memory to MEMORY.md (2,200 characters) and USER.md (1,375 characters); the agent writes its own memory and prefetches each round.

Unified Benchmark – OmniMemEval

OmniMemEval runs 14 mainstream memory products on a single pipeline with identical model configuration, prompts, and judge. MemOS achieves the highest scores across the benchmark.

User Memory Evaluation

In the user‑memory track (long‑term memory for individual users), the cloud version of MemOS records:

LoCoMo = 92.34

LongMemEval = 93.40

These values reach industry‑SOTA levels. The Context Tokens metric shows that MemOS attains better performance with fewer injected tokens, resulting in lower API cost, faster response time, and reduced hallucination risk.

Agent Memory Evaluation

MemOS Local Plugin 2.0 adds local long‑term memory to OpenClaw or Hermes. AgentBench evaluates five task domains:

BrowseComp‑Plus – information retrieval

OmniMath – mathematical reasoning

SWE‑Bench – software engineering

LiveCodeBench – code implementation

GDPVal – knowledge work

Results for OpenClaw show MemOS ranking first in four of the five tasks (BrowseComp‑Plus, OmniMath, SWE‑Bench, GDPVal). The average task‑completion rate rises from 36.63 % to 50.87 % , a structural capability boost that cannot be achieved by prompt tuning alone.

Hermes‑specific results:

Baseline OmniMath = 62.00, MemOS = 72.67

Baseline SWE‑Bench = 37.18, MemOS = 52.56

The >15‑point increase on SWE‑Bench indicates that long‑term memory markedly improves bug‑fixing ability.

Installation of MemOS Local Plugin 2.0

Install with a single command:

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

The script automatically detects existing OpenClaw or Hermes installations, deploys the plugin to ~/.hermes/plugins/, generates config.yaml, and launches the Viewer.

The Viewer runs at http://127.0.0.1:18800, displaying all memory layers. The first launch requires setting a password.

Within the Viewer configuration panel, Hermes’s native memory can be imported and examined in the workspace.

Enabling the “memory self‑evolution” feature allows automatic recording of tasks, experiences, environment cognition, and skills. The feature can be configured to use the Qwen‑3.8‑max‑preview model.

Memory Growth and Skill Extraction

MemOS consolidates scattered memories into a single panel and lets memory continuously grow: repeated tasks become records, validated methods become experience, project and environment understanding matures into cognition, and mature knowledge crystallizes into skills.

Key Visuals