Memory

Agents have persistent memory that survives across conversations. No vector database, no RAG — just two plain markdown files that work.

Two files

• HISTORY.md — append-only conversation log. Timestamped summaries of every conversation. Auto-appended after each session.
• NOTES.md — agent-curated knowledge. Facts, preferences, learned context. The agent writes to this using the memory_write tool.

Context integration

On every agent loop iteration, the context builder:

• Grep-searches HISTORY.md for terms relevant to the current message
• Includes the full NOTES.md contents

This gives the agent relevant historical context without flooding the context window with the entire conversation history.

memory_read tool

Agents can explicitly search their memory using memory_read. Pass a query string and it returns matching lines from both HISTORY.md and NOTES.md.

memory_write tool

Agents curate their own knowledge by writing to NOTES.md. Typical entries:

User prefers responses in Norwegian.
Budget for Q1 project: 500,000 NOK.
Primary competitor: Acme Corp (acme.no).

Storage

Memory files are stored in R2, scoped per organization. Each org's workspace at orgs/{orgId}/workspace/memory/ contains both files. You can view and edit NOTES.md directly from the dashboard.