OPENCLAW MEMORY:

LEVELS UNLOCKED

Barron Roth · 2026

PRESS START

→ or SPACE to advance

MY STORY

Deep in a Cubby coding session one weekend — building a new feature, making architectural decisions, debugging edge cases. Real flow state.

My AI assistant remembered every decision, every constraint, every thread. It felt like having a genius collaborator.

Monday through Friday — other things. Come back next weekend...

"Hey, let's pick up where we left off on Cubby."

CONTEXT LOST

Architecture decisions, debugging context, the whole thread — vaporized.

✈️

PREFERENCE AMNESIA

"Every time I asked for flights: nonstop, United, no red-eyes. Every. Single. Time. Like onboarding a new hire every session."

🎵

CONTRADICTED DECISION

"I spent 30 minutes picking a wedding DJ. A week later my agent pitched me the one I explicitly rejected."

💡

"The system didn't get worse — my bar got higher. The more I relied on it, the more the cracks showed."

HOW MEMORY

ACTUALLY WORKS

"Most people think memory is one thing.
It's actually four different systems."

LAYER 1: BOOTSTRAP FILES

✅ LOADED FROM DISK EVERY TURN — SURVIVE COMPACTION

📕 MEMORY.md

Curated long-term memory. Decisions, project status, vendor contacts, lessons learned. Main session only.

📜 AGENTS.md

Operating instructions. "Search memory before acting." Rules, workflows, safety constraints.

📅 memory/YYYY-MM-DD.md

Daily logs. Append-only. Today + yesterday loaded at session start. Raw context, not curated.

🔧 SOUL / USER / TOOLS.md

Personality, human profile, tool configs. Stable reference files. Also loaded every turn.

THE 4 LAYERS

✅

LAYER 1: BOOTSTRAP FILES

SOUL.md, AGENTS.md, MEMORY.md — loaded from disk every turn

❌

LAYER 2: SESSION TRANSCRIPT

Compaction summarizes it. Detail evaporates.

⚠️

LAYER 3: PRE-COMPACTION FLUSH

Safety net — but most people never check if it's working

⚠️

LAYER 4: RETRIEVAL INDEX

Stock: auto-selects OpenAI → Gemini → Voyage → local embeddings. No API key = no search. Only indexes MEMORY.md by default.

NUCLEAR COMPACTION

💣

▼

ONE-WAY DOOR. DETAIL GONE FOREVER.

"If it's not written to a file,
it doesn't exist."

That magic I felt when I first set up OpenClaw — zapped away.
It's not magic. It still has limitations. Memory might be the biggest one.

LEVEL UP!

LEVEL 2

CONTEXT ISOLATION

THE PROBLEM

One context window doing everything.

💒 📱 💰 ✈️ 🏋️ 📧

Wedding vendors polluting Cubby coding sessions. Tax questions muddying travel planning.

THE UPGRADE: TOPIC ISOLATION

💒

WEDDING TOPIC

Vendors, planning, budget

📱

CUBBY TOPIC

Code, architecture, bugs

💰

FINANCE TOPIC

Taxes, investments, bills

Each topic = its own session with its own context window

PER-TOPIC SYSTEM PROMPTS

📱 cubby → reads cubby.md

💒 wedding → reads wedding.md

🍽️ maître d → reads maitred.md

✈️ travel → reads travel/*.md

💳 credit cards → reads credit-cards.md

💰 finances → reads budget.md

channels.telegram.groups.<id>.topics.<threadId>.systemPrompt

LEVEL UP!

LEVEL 3

RETRIEVAL (QMD)

WHAT IS QMD?

🔍📜📜📜

On-device semantic search over session transcripts

638

transcripts

68K

vectors

ZERO

cloud

THE HIDDEN SWITCH

sessions.enabled

ON

OFF by default. Without this, QMD only indexes MEMORY.md.
You miss 99% of searchable history.

RAG-TYPE FLAWS

✨📜 ✨📜 📜 📜 ❌📜

🎲

Probabilistic, not deterministic. Embedding similarity isn't the same as understanding.

📡

Signal-to-noise at 600+ transcripts. Mentions in passing vs. deep discussions.

🤷

Reactive — agent must KNOW it forgot. If it doesn't realize it's missing context, it guesses confidently.

QMD = "I forgot, let me look it up."

Cross-session recall unlocked. But it's a search engine, not a memory system.

LEVEL UP!

LEVEL 4

LOSSLESS CONTEXT MANAGEMENT

THE KEY DISTINCTION

QMD

"I forgot, let me
look it up."

REACTIVE

LCM

"I never fully
forget."

PROACTIVE

STOCK COMPACTION

💣

ONE-WAY DOOR

LCM

▼

BRANCH

▼

CANOPY ✦

TWO-WAY DOOR

THE DAG

Condensed Summaries (Canopy)

Navigation layer

▲

Condensed Branch

Meso-compaction

▲

Leaf Summary

Micro-compaction

▲

Raw Messages (never deleted)

Ground truth

⬇️ lcm_expand — drill back down to recover detail anytime

CHARACTER STATS

session.reset.mode

idle (7 days)

Sessions survive restarts

freshTailCount

128

Protect last 128 messages

contextThreshold

0.75

Compact at 75% full

incrementalMaxDepth

-1

Unlimited DAG depth

QMD + LCM TOGETHER

LCM

Within-session
memory

DAG compaction
Micro-summaries
lcm_expand

BOTH ACTIVE
Complementary,
not competing

QMD

Cross-session
memory

638 transcripts
Semantic search
memory_search

"Nothing is ever deleted
from the database.

The summaries are a navigation layer,
not a replacement."

STAGE CLEAR!

THREE RULES

🏆

RULE 1

Put durable rules in files, not chat. MEMORY.md, AGENTS.md — these survive everything. Chat instructions vanish.

🏆

RULE 2

Check that safety nets work. Memory flush, embed intervals, session reset settings. Most people never verify.

🏆

RULE 3

Make retrieval mandatory — add "search memory before acting" to AGENTS.md. Without it, the agent guesses instead of checking.

"The system keeps evolving. My job is to keep raising the bar."

alfred.barronroth.com/guide-to-alfred

GAME OVER

THANK YOU