Core Concepts

OpenClaw relies on three core concepts—memory, context, and soul (agent identity)—so the assistant can behave consistently across conversations and stay within model limits.

Memory is how OpenClaw stores and recalls information over time. Memories can be file-based (e.g. in your workspace), embedded in a vector store for semantic search, or surfaced via daily notes and integrations like Obsidian. The agent is instructed to read and update memory so it can refer to past decisions, preferences, and facts. Understanding memory helps you control what the assistant “knows” and where that data lives.

Context is the conversation window sent to the model on each turn: recent messages, relevant memory, and system prompts. Token limits mean context cannot grow forever; OpenClaw compacts or summarizes when needed so long threads stay usable. Configuring context (e.g. how much history to include, when to compact) affects reply quality and cost.

Soul (agent identity) is who the agent is—defined in workspace files like SOUL.md and AGENTS.md. Research shows generic labels ("You are a helpful assistant") add no measurable benefit; detailed, specific souls do. Put identity first in the system prompt; describe roles experientially or use two-stage soul design (let the model draft its expert identity); use multi-soul review for high-stakes work. See Soul & Agent Identity for the full picture.

Explore the sections below for details on each system.