Friday Talk
OpenClaw: Forget About It?
What AI agents actually remember - and what they don't
Yan Gonzalez - True Webmaster
March 20, 2026
The Hook
"I corrected my agent. It said 'I understand.' Next day, same mistake. I corrected it again. 'I understand.' This went on for weeks."
The corrections weren't sticking. And I couldn't figure out why.
The Illusion
"Your AI agent says 'I remember that project.' It doesn't. It's predicting what a helpful assistant would say next."
- ChatGPT's "memory" is a list of notes - not understanding
- Most agents start every session from a blank page
- Real memory requires engineering - not just a bigger context window
The Capability Ladder
What your agent can actually do - and where the illusion begins.
Most agency owners think they're at L4 or L5 after a few corrections. Reality: you're between L2 and L3.
The Semantic Gap
Why your corrections don't stick. Two systems. One gap.
SOUL.md - personality, rules
AGENTS.md - operating procedures
MEMORY.md - your corrections
Prompt context only. Influences reasoning. Not enforced.
Tool policies - deny lists
Exec approvals - human gates
Sandbox - isolation
Deterministic. Code-level. Cannot be bypassed.
Your rules exist in the brain. Execution happens in the hands. The hands don't read the brain.
Hooks
Code that bridges the gap. Deterministic. Not optional.
Injects remembered facts into the prompt before the model sees it. The agent doesn't choose to remember. The code forces memory into context.
Injects relevant search results automatically. The agent receives enriched context without asking for it.
Logs every action the agent takes. The agent can't skip being observed. Every tool call is recorded.
Intercepts outgoing messages before they reach the user. Filter, modify, or block at the code level.
The brain can't read the hands. But hooks can wire them together.
Memory - The OpenClaw Definition
How OpenClaw and its ecosystem define agent memory today.
This is how most people understand it. This is what the docs describe. And this is what most memory plugins in the ecosystem are trying to improve.
Memory Framework
Cognitive functions mapped to plugins - seriously.
| Cognitive Function | Human Equivalent | Plugin / System | What It Does |
|---|---|---|---|
| Working Memory | What you're actively thinking about | Context Window | Holds active context - limited size is the root constraint |
| Context Management | What gets into working memory | LCM | Compresses and reconstructs context - prevents loss from compaction |
| Episodic Memory | Experiences / diary | Core Logs | Stores raw conversations - pure storage, no understanding |
| Episodic Retrieval | Remembering past events | QMD | Searches and retrieves from logs - diary-based, not knowledge |
| Semantic Memory | Facts, knowledge, understanding | TrueMem (Graph) | Stores structured facts + relationships - this is the gap |
| Consolidation | Experience → knowledge | TrueMem Librarian | Extracts facts from conversations into the graph |
| Retrieval (Semantic) | Recall of relevant knowledge | TrueMem Retrieval | Injects relevant facts into context alongside QMD |
| Forgetting | Removing outdated knowledge | Temporal Logic | Expires old facts, updates graph - time-aware memory |
| Procedural | Skills / how to do things | OpenClaw Skills | Executes workflows and actions - independent from knowledge |
LCM - Context Management
Layer 1: How the agent keeps its thread without losing information.
The problem
- AI context windows have hard limits (200K tokens)
- Long conversations get silently truncated - old context is gone
- The agent forgets what you said 30 minutes ago
What LCM does
- Lossless compaction - summarize old turns without losing detail
- DAG-based summaries - drill back into any compacted section
- Keeps the important context, compresses the noise
- The agent's conversation thread never truly breaks
Without this, every long session ends with the agent asking "what were we working on?" LCM prevents that.
QMD - Episodic Memory
Layer 2: What happened today? What did we talk about last week?
What it does
- Markdown files on disk - daily notes, decisions, learnings
- Semantic search across all memory files
- The agent's journal - raw, unstructured, fast to write
- Survives restarts, context resets, even full redeployments
How it works
MEMORY.md- curated long-term contextmemory/YYYY-MM-DD.md- daily logsLEARNINGS.md- rules from mistakesmemory_search→ semantic recall across all files
Think of it as the agent writing in a notebook every day. Simple, durable, searchable.
TrueMem - The Stack
Four layers working together. This is the ecosystem behind semantic memory.
Not a single tool - an ecosystem. Each layer does something the others can't.
TrueMem - How Memories Get Written
Consolidation: turning conversations into permanent knowledge.
This runs every 5 hours automatically. Per-agent routing means each agent's facts stay private - no cross-contamination.
TrueMem - How Memories Come Back
Retrieval: every message triggers a search. Here's what happens in ~3 seconds.
TrueMem - What Makes It Different
Features that vector search doesn't have and probably never will.
Facts expire. Client moved? The old address is invalidated, not deleted. The graph knows what was true vs what is true.
Different names, same thing. Graphiti resolves "Acme", "Acme Co", "Acme Roofing" into one entity automatically.
Each agent's facts stay in their own group. No cross-contamination. Your SEO agent doesn't see your dev agent's debugging.
Filters out 50% of unnecessary searches - system messages, heartbeats, and internal noise never hit the database.
Weekly automated cleanup removes stale connections, disconnected records, and junk data. The graph stays healthy on its own.
See exactly what memory your agent used for every response. You can't optimize what you can't see.
Cost: ~$0.86/day for the full stack. Vector search is cheaper but misses all of this.
You Can Have This
TrueMem is free. The plugin, the librarian, the cleanup module, the config - all of it.
It's a private repo. Give me your email and I'll add you.
No strings. No paywall. Just share your email after the talk.
📦 TrueMem Plugin (auto-recall, skip gate, cache, noise penalty)
📚 Librarian (batch extraction, entity resolution, group routing)
🧹 Cleanup Module (automated weekly graph hygiene)
⚙️ Config + Documentation
That's what it takes to make one agent remember.
Four layers. Nine cognitive functions. Seven LLM calls per fact. Weekly cleanup. Temporal decay. Entity resolution.
Now let's zoom out.
The Iceberg
What users see is just the tip.
Claude
What Users See
- Context Management
- Tool Calling
- Memory Systems
- Knowledge Graphs
- Embeddings
- Retrieval
- Orchestration
- Security
- Infrastructure
The AI Control Ladder
Where are you? Every step up gives you more power - and more responsibility.
The System > The Model
These companies aren't just building better models - they're engineering what goes around them.
Claude Code (Anthropic)
- Terminal-first - bash is the harness
- Sandboxed, permission-gated
- Progress files for memory across sessions
- 84% fewer approval prompts (Anthropic, 2026)
Codex (OpenAI)
- Sealed containers - repo is source of truth
- PR-based workflow, mechanical enforcement
- Linter-as-remediation pattern
- 1,500 PRs merged, zero manual lines (OpenAI, 2026)
Cursor (Anysphere)
- Planner → Worker → Judge hierarchy
- Fresh restarts, no stale context
- 1,000 commits/hour at scale
- 10M tool calls per week (Cursor, 2026)
Billions invested - not in bigger models, but in better systems around them.
Before You Start
Honest advice from someone who's been through it.
Every tool, every plugin, every deployment runs through Git. If you don't know GitHub, you're locked out of the ecosystem. Rewatch Brian Kato's Mastermind session.
DNS, tunnels, Pages, Workers, security. Cloudflare is the infrastructure layer for everything you deploy. This site runs on Cloudflare Pages.
You can't use these tools if you can't read the language. The AI Nerdccionary on this site has 217 terms across 13 categories. Speaking software is the price of admission.
Don't install everything on day one. Start simple, learn the basics, add complexity only when you need it.
Google's open-source terminal agent. Free (1,000 requests/day). Has GEMINI.md persistence, skills, MCP support. A solid free starting point before you commit money.
More harness than OpenClaw. More general-purpose than Codex. No middleware, no ToS risk. $100-200/mo on Anthropic stretches far for agency work.
Claude Code GA (May '25). Web (Oct '25). Co-Work Mac (Jan '26). Co-Work Windows (Feb 11). Remote Control (Feb 25). The gap between platforms is closing every month.
Uses other people's models. Anthropic put OpenClaw against their ToS. Get banned and your operation goes dumb overnight. Understand the risk before you build on it.
Launched Feb 2026, $200/mo. Uses Claude, GPT, Gemini behind the scenes. Impressive but middleware. Will get squeezed when vendors ship their own versions.
What You Get Today
On This Site
- The White Paper
- 🎙️ The Deep Dive (audio)
- 🎙️ The Debate (audio)
- Slides + Memory Framework
- AI Nerdccionary + 3 Memory Guides
- Telegram + CLI cheat sheets
Your Next Steps
- 1. Master GitHub + Cloudflare
- 2. Start with Claude Code or Gemini CLI
- 3. Learn the memory framework
- 4. Read the paper. Understand L1-L5.
Password: memory2026