📝 MEMORY.md — The Readable Layer
Hermes’ first layer of memory is a simple text file called MEMORY.mdIt’s plain Markdown: you can open it in any editor, read it, and edit it. It lives on your laptop — not on a third-party server. Think of it as the little notebook where the agent records what matters about you.
📒 Why Markdown and not a “magic” database
Plain text has three advantages that a closed system doesn’t offer:
- •Transparent: you can see exactly what it “knows” about you.
- •Editable: corrects an incorrect fact by opening the file.
- •Portable: is just a file; back it up, copy it, version it.
What MEMORY.md looks like (illustrative)
# MEMORY.md ## Pessoa - Nome: Nei - Aniversário: 14 de março - Fuso: America/Sao_Paulo (BRT) ## Preferências - Gosta de respostas diretas, sem enrolação - Trabalha com automação (n8n, Make) ## Projetos ativos - Curso "Hermes 21C" — entregar trilha 2
🔍 Full-text search — recall everything
The second layer is a full-text search in SQLite about EVERY message from EVERY session. While MEMORY.md keeps the "curated summary," this layer stores the raw history and lets you find any phrase by keyword or date.
💬 Real-world recall example
You can literally ask:
"Give me an example of a question I asked in May 19."
The agent searches the indexed history and returns the exact message — something impossible with only the curated MEMORY.md.
💡 Practical tip
The two layers complement each other: MEMORY.md is what it always has in mind; full-text search is what it will fish when you need an old detail.
⏳ Composes silently
Memory doesn’t get good on the first day. It composes in the background as you use it, and takes about 2 weeks of real usage for the agent to feel “magical” — knowing your patterns, your vocabulary, your priorities.
Day 1–3 — almost like a chatbot
Little accumulated context
It doesn’t know you yet. The answers are generic; you have to repeat context.
Week 1 — starts to remember
Patterns emerging
It already avoids asking about things you’ve told it; it makes suggestions based on what it knows.
~2 weeks — “magical”
Composite memory
Anticipates your needs, connects projects, and talks like someone who’s been with you for a long time.
💡 Practical tip
Don’t give up in the first few days. The feeling that “it’s not worth it” is normal before the memory takes shape. Use it for real for 2 weeks before judging.
🔗 Supercharge — boost memory
You can connect external sources so Hermes can "know" much more: Obsidian (your notes), Pinecone (vector memory), the memory of the Claude Code and even apps on your computer.
✓ Good sources to connect
- ✓Obsidian — your second brain for notes
- ✓Pinecone — semantic recall at scale
- ✓Claude Code memory — projects and decisions
✗ Precautions when boosting
- ✗Don’t dump everything in at once (it becomes noise)
- ✗Don’t connect sensitive data unless necessary
- ✗Don’t forget: more context = more tokens per turn
📊 The effect of boosting
- More sources → more specific and personalized answers
- Semantic recall → finds by meaning, not just exact words
- Unified memory → connects what you say across different channels
🌙 It "dreams" at night
During idle periods, Hermes processes what it has learned and returns proactively with suggestions. It’s the difference between passive memory (which only responds when asked) and active memory (which connects the dots on its own).
🌙 What “dreaming” means in practice
It's not mysticism — it's idle processing that revisits the history and generates insights:
- •"I noticed you’ve put off X three times—want me to schedule it?"
- •"That project you mentioned has a deadline tomorrow."
- •"You mentioned working out; want a recurring reminder?"
💡 Practical tip
The richer the memory, the better the "dreams." This proactive behavior connects to the heartbeat/cron in Track 3 — the agent that comes to you.
🌐 The "everything system"
The core philosophy: without information about what you do, there’s no way to give precise advice. Memory is what turns a generic model into an assistant that understands your context — the "everything system".
✗ No memory
- ✗Generic, “by-the-book” advice
- ✗You repeat context every time
- ✗Each session starts from scratch
✓ With memory
- ✓Advice tailored to your situation
- ✓Pick up where you left off
- ✓Connects projects and decisions over time
🌐 Why this wraps up the capabilities track
Soul (2.2) provides personality, integrations (2.3) provide reach, actions (2.4) provide arms — but memory is what provides continuityWithout it, all the other capabilities restart with every conversation.
📌 Module Summary
Next Module:
2.2 — Soul.md: the character’s bible