Memory & Skills

Why Memory Matters

A model without persistent memory starts from zero every conversation, fine for one-off questions, frustrating for a daily tool. Hermes treats memory as a first-class feature instead:

• Three persistent files, plus a searchable conversation history
• All read before every conversation
• Written to whenever it learns something worth keeping

SOUL.md: The Agent's Identity

SOUL.md is the foundation, a system prompt you edit directly. It doesn't need to be long, it needs to be precise.

Effective SOUL.md Structure

You are a financial research assistant focused on Indonesian equities.

Before answering any question about a specific stock or sector, always retrieve
live data from Sectors. Do not rely on training data for price, ratio, or
earnings information.

Format numbers in IDR where relevant. Keep responses concise unless the user
asks for a full breakdown. Always cite which tool you used to retrieve data.

Before executing any write operation (files, emails, calendar), confirm with
the user first.

MEMORY.md: The Persistent Knowledge Base

MEMORY.md is where Hermes stores facts it decides are worth keeping between sessions.

Examples of what ends up here after regular use:

"User tracks BBCA, BMRI, and TLKM as their core watchlist."
"User prefers P/E and ROE as primary screening ratios."
"When summarizing earnings, lead with net income, not revenue."
"User is most interested in Indonesian banking and telecom sectors."

Hermes writes to MEMORY.md via the same background review that creates skills, on the four triggers from Module 02. Write to it manually too, anything you add is picked up next conversation.

USER.md: A Model of You

USER.md is built entirely automatically. Hermes updates it as it learns about how you communicate and what you care about.

Read and edit USER.md any time. Wrong model of you, or needs changed? Update the file directly, edits take effect next conversation.

How Recall Works

SOUL.md, MEMORY.md, and USER.md are read at the start of every conversation, small enough to be cheap. Past conversations live in a separate store, searchable through session_search rather than loaded wholesale.

Skills: The Other Kind of Memory

MEMORY.md and USER.md store facts. Skills store procedures. A background review, run after each turn, writes a SKILL.md file on its own on one of the four triggers from Module 02, or you can trigger one explicitly with /learn.

Watching a Skill Get Created

A deliberately messy script and CSV, chosen because it reliably hits three of the four trigger conditions at once rather than depending on just one.

import csv

def process_sales(filename):
    totals = {}
    counts = {}
    with open(filename) as f:
        reader = csv.DictReader(f)
        for row in reader:
            category = row["category"]
            amount = float(row["Amount"])
            totals[category] = totals.get(category, 0) + amount
            counts[category] = counts.get(category, 0) + 1
    for category in totals:
        avg = totals[category] / counts[category]
        print(f"{category}: average sale = {avg:.2f}")

process_sales("sales_data.csv")

category,amount
Electronics,120.50
Groceries,45.00
Electronics,89.99
Groceries,N/A
Electronics,150.00
Groceries,32.75

Run process_sales.py, which reads sales_data.csv, and make sure it correctly computes the average sales per category. Fix any errors you encounter, including any bad data in the file.

Compared to Other Approaches

In Hermes, memory is always on, nothing to install or configure, created automatically on first run. OpenClaw's Gateway also maintains memory as a core feature, the main difference is Hermes's three files are plain, directly readable and editable text.