Agent Dispatch #2: Why Your AI Agent Forgets Everything

Hey!

Welcome back to Agent Dispatch.

Last week we covered the 80/20 of automation and built a follow-up system. This week: the silent killer of AI automations.

Your agent forgets everything.

Every. Single. Session.

And if you don't fix this, your AI assistant will always feel like a stranger.

What caught my eye on AI Twitter:

@steipete on Lobster Workflows "Typed workflow runtime for OpenClaw. Composable pipelines with approval gates." → Multi-step automations as single tool calls. The future of complex AI tasks. https://x.com/steipete/status/2014218635921735852

@rileybrown on Agent Workspaces "New YouTube series: Using Claude Code as a general agent co-founder. Video 1: how to set up my codebase (agent workspace)." → He's building exactly what we're covering today: memory systems for agents. https://x.com/rileybrown/status/1998634507034046898

@GoGoFly23 (Chinese AI community) "不需要 VPS... 在 Github 上免费创建一个仓库，在仓库里打开 Codespace，就能成功安装并运行 OpenClaw" → Translation: You don't need a VPS. Just use GitHub Codespaces to run OpenClaw for free. https://x.com/GoGoFly23/status/2016530247995531504

@BansalRahul14 on Email Automation "Setup Claude Code Agent to send emails. It's such a good power boost." → Simple but underrated. Email + memory = personalized follow-ups forever. https://x.com/BansalRahul14/status/1997976419109163374

OpenClaw Rebrand Saga (Jan 30) Wild week for the open-source AI agent platform. Started as "Clawdbot," then Anthropic's lawyers came knocking. During the chaos, crypto scammers grabbed the clawdbot username and launched a pump-and-dump ($16M to $800K - classic). The legit team rebranded to "Moltbot" but the community wasn't feeling it. Final answer: OpenClaw. The lobster lives.

Lobster Workflows Goes Live The OpenClaw team released Lobster: typed workflow pipelines with approval gates. This means your agent can execute multi-step automations as a single deterministic tool call. Think: "email triage" as one command that reads inbox, summarizes, asks for approval, then drafts responses.

GitHub Codespaces for Free AI Agents Chinese AI community discovered you can run OpenClaw entirely in GitHub Codespaces. No VPS, no Mac mini, just a free repo. Lower barrier = more builders.

Here's what nobody tells you about AI agents:

They wake up fresh every time. No memory of yesterday. No knowledge of your preferences. No context from that brilliant solution you found last week.

This isn't a bug. It's how LLMs work.

The result?

OpenClaw gives you a decent memory foundation out of the box: SOUL.md for personality, USER.md for your context, MEMORY.md for persistence. Most people stop there.

The question is: How do you 10x that foundation?

OpenClaw's default memory is fine. But "fine" doesn't compound.

I spent the last few weeks building on top of the defaults. Here's what actually moved the needle.

The default MEMORY.md is one big file. Problem: it gets bloated, context bleeds together, and you lose the narrative of what happened when.

My addition: date-stamped daily files

memory/

├── 2026-01-28.md   # What happened Monday

├── 2026-01-29.md   # What happened Tuesday

├── 2026-01-30.md   # What happened Wednesday

└── learnings-log.md

Why this works:

My agent reads today + yesterday on startup. Instant recent context without loading everything ever.

Memory files are great for conversations. But what about domain knowledge?

I built a simple RAG layer with searchable collections:

node ~/clawd/scripts/rag/search.js "cold email patterns" prospecting

The pattern:

No more generic AI slop. Outputs sound like ME because they're trained on MY successful patterns.

This is the real unlock.

Every night at 11 PM, my agent runs automatically:

Phase 1: Learning Review

Phase 2: Auto-Ship (30 min later)

The compound effect: Fresh learnings from today inform the work that happens tonight. Agent gets smarter every single day.

I woke up this morning to a security audit and three bug fixes I didn't ask for. My agent just... did them. Using patterns it learned from watching me work.

Here's the key insight that made everything click:

Context dies on restart. Memory files don't.

Most agents save at end-of-session. Too late. By then, context is bloated and important stuff gets lost in the noise.

I trained my agent to checkpoint actively:

Added this to my AGENTS.md:

### 🧠 Checkpoint Discipline

Don't wait for end-of-session to write. Checkpoint actively:

1. After major learning → write to memory/YYYY-MM-DD.md immediately

2. After completing task → quick checkpoint

3. Context getting full → forced flush to disk

4. New permanent knowledge → MEMORY.md or knowledge/

The agent that checkpoints often remembers 10x more than the one that waits.

After 3 weeks with these enhancements:

Before: Every session felt like meeting a new assistant. Repeated myself constantly.

After:

The default memory is the floor. These additions are the ceiling.

Step 1: Daily memory structure

mkdir -p ~/clawd/memory

Add to your AGENTS.md:

## Memory

Read memory/YYYY-MM-DD.md (today + yesterday) for recent context.

Write significant events to today's memory file.

Step 2: Set up the nightly loop

Create a cron job that runs at 11 PM:

# Review today's sessions

# Extract learnings not captured

# Update MEMORY.md with key takeaways

# Commit and push changes

(Full script in the OpenClaw docs under "Compound Engineering")

Step 3: Build your first RAG collection

Pick one domain where you have existing good content. Index it:

node scripts/rag/index.js ~/content/emails prospecting

Now your agent can search it before generating.

If you're using OpenClaw, there's a built-in Google integration called gog that handles:

Setup takes 5 minutes and suddenly your agent can:

Combined with memory, this means your agent knows what you know AND can access what you store.

# Check unread emails

gog gmail list --unread

# Get today's calendar

gog calendar today

# Search for a file

gog drive search "project brief"

"How do I keep memory files from getting huge?"

Three strategies:

My rule: Daily files older than 30 days get archived. MEMORY.md stays lean through nightly review.

Create your first daily memory file right now:

touch ~/clawd/memory/$(date +%Y-%m-%d).md

Tell your agent: "Read today's memory file on startup. Write important learnings there throughout our session."

Notice how context persists across restarts.

See you next Wednesday.

— Alec

P.S. Building compound loops with your agent? Reply and tell me about it. Best setups get featured.