Guide

Memo.st is a memory layer for AI agents. It stores facts that should survive a single chat session, then retrieves the right context when an agent needs to act again.

Why memory exists

Agents often repeat work because the useful context from previous sessions is trapped in transcripts, tickets, emails, or tool logs. Memo.st turns that context into explicit memory records with tenant boundaries, ownership, metadata, and retrieval behavior.

What gets stored

A memory is a durable statement about a subject:

Acme prefers Slack escalation for P0 incidents.

Memories can represent preferences, facts, events, summaries, decisions, constraints, or relationships. They should be short enough to retrieve and specific enough for an agent to use without rereading the whole source document.

How records are scoped

Every memory belongs to an organization and is grouped by subject, agent, and namespace.

organization -> subject -> namespace -> memory
                         -> agent

This shape keeps retrieval predictable:

• Organization boundaries isolate customer data.
• Subject IDs identify who or what the memory is about.
• Agent IDs record which system created or owns the memory.
• Namespaces separate workflows like support, sales, and research.

How retrieval works

The API supports two retrieval paths today:

• D1 fallback retrieval returns recent memories for an organization and optional namespace.
• Vectorize retrieval uses caller-provided query embeddings to find semantically similar memories.

The long-term direction is hybrid retrieval: exact filters, semantic search, and graph-aware expansion.