Memory

Memory in OpenWalrus is file-per-entry. Each memory is a plain markdown file you can read and edit yourself. Recall uses BM25 ranking with zero external dependencies. No vector databases, no embedding services.

How it works

• Each memory is a markdown file in the agent's memory directory
• Relevant memories are auto-recalled before each conversation turn using BM25 ranking
• MEMORY.md is your agent's curated overview — always loaded into context
• Walrus.md is your agent's editable identity
• Agents decide when to store and retrieve — the framework doesn't impose a strategy

Design philosophy

Traditional agent frameworks hardcode a memory layer: a vector database, a journal system, a knowledge graph. OpenWalrus takes a different path. Memory is a folder of markdown files. You can read them, edit them, delete them. Your agent does the same.

This means:

• No infrastructure — no databases to run, no embedding pipelines to maintain
• Human-readable — every memory is a markdown file you can inspect
• Agents curate their own knowledge — the LLM decides what to remember and what to forget
• Old formats migrate automatically — upgrading preserves your agent's memory

What's next

• Skills — prompt-level behaviors that can guide memory patterns
• Context compaction — how agents preserve knowledge across long sessions
• Extensions — runtime extensions that can augment memory capabilities