AI Agents Now Have Two Kinds of Memory — and Mixing Them Up Is the Core Architecture Mistake

Developers building AI agents have long wrestled with the same problem: how do you give an agent memory that persists? Context windows end, sessions change, and the agent "forgets" everything. A clearer model is now emerging — not one type of memory, but two distinct systems with different functions.

This week @garrytan (Garry Tan, Y Combinator president) reshared an insight from the Japanese developer community that articulated this distinction precisely:

OpenClaw memory is the "agent's operational memory." This stores preferences, working rules, and current project context. Think of it like RAM — fast, contextual, operational.

GBrain is the "external brain of world knowledge." This stores information about people, companies, meetings, and ideas. Semantic search over your notes, documents, and contact base. Think of it as long-term memory with intelligent search.

Consider two different questions an AI agent might receive:

Question 1: "How do I want you to communicate with clients?" — this works through operational memory. The agent needs to know your communication style, rules, and preferences.

Question 2: "What did we discuss with Baltic Inc. in our last meeting?" — this works through world knowledge. The agent needs to search for specific information about a specific entity.

Both questions look similar, but they require different mechanisms. Conflate them and you get either a slow system (if you put everything in semantic search) or an "amnesiac" system (if everything lives in context and runs out).

Per the model shared by @garrytan, both systems operate through the same interface — CLI commands and an MCP server. An agent can:

1. Check its operational rules (OpenClaw memory) before each action
2. Search for contextual information (GBrain) when specific facts are needed
3. Update both systems after each session

A concrete real-world example: developer @SajiBhai011 shared that they indexed 116 pages of Obsidian notes, split into 479 chunks, through their GBrain setup. Now they can query and get answers from their own note base — not from the internet, not from model training data, but from personal documents.

As @kakumaeda put it (retweeted by @garrytan): "OpenClaw memory is 'agent operational memory', GBrain is 'external brain of world knowledge'... will enable cross-agent operation via CLI and MCP server." — X

This architecture matters for anyone building or planning AI agents for business processes. In practice:

• Operational memory: communication style, product rules, escalation procedures

• World knowledge: specific customer history, prior conversations, agreement details

• Operational memory: sales playbook, objection responses

• World knowledge: specific contacts, company profiles, deal history

• Operational memory: personal preferences, working style

• World knowledge: project documents, meeting notes, idea repository

The dual-memory architecture isn't a technical curiosity — it's a practical answer to a real problem: AI agents need to know both how to behave (rules) and what to talk about (facts). Conflating these two functions is the primary mistake in building long-lived agents.

European businesses adopting AI agents for core processes will need to design for both. The good news: the tooling is maturing fast, and the architectural pattern is becoming clear.

WebEdge.dev implements multi-agent systems for businesses in Lithuania and across Europe, including memory architecture following these principles.

---

The agent's operational memory system — stores working rules, preferences, and current context. Operates via CLI and MCP server.

A semantic search system over a personal knowledge base — notes, documents, contact information. Garry Tan is building it as a separate layer on top of Claude/OpenAI models.

For long-running business agents — yes. Operational memory ensures consistent behavior. World knowledge ensures factual accuracy.

Yes — developers are already indexing their Obsidian notes through GBrain and building agents that search their personal knowledge base for answers.