Core Concepts

Memory System

GoClaw uses a two-tier memory system based on plain Markdown files. No vector databases, no embeddings, no SQLite — just files you can read and edit in any text editor.

Daily logs

Every conversation turn is appended to ~/.goclaw/memory/YYYY-MM-DD.md. GoClaw loads the current and previous day's logs into every LLM context window automatically.

~/.goclaw/memory/2026-02-19.md
markdown

Long-term memory

Facts that should persist indefinitely live in ~/.goclaw/memory/longterm.md. The agent populates this via the built-in remember skill:

conversation example
text
~/.goclaw/memory/longterm.md
markdown
â„šī¸
The number of past daily logs loaded into context is configurable via memory.days_back in config.yaml. Default is 2 (today + yesterday).

Why not a vector database?

Vector databases are powerful but add significant operational complexity — they require embeddings, a running service, and specialized query logic. For a personal assistant that runs on a single VPS or Raspberry Pi, this overhead isn't worth it.

Plain Markdown files are human-readable, git-trackable, editable in any text editor, and trivially fast to load. The current approach handles most personal use cases comfortably within a 200k+ token context window.