Skip to main content

Memories

Memories are durable facts that survive across conversations. They come in two scopes backed by one store:

  • Agent memories — things an agent learned about a specific user: facts, preferences, context, and standing instructions. Scoped to (agent, user), so each agent remembers its own relationship with each person.
  • Workspace memories — the workspace's operating brain: guidance, preferences, decisions, learnings, and facts that the planning and strategist layers consult when deciding what the workspace should do next.

Open Memories in the sidebar to browse, add, edit, archive, or delete them, filtered per agent.

Where Memories Come From

  1. Manual — you write them yourself. Manual entries are treated as the most authoritative source.
  2. Conversation extraction — an LLM pass over a conversation pulls out memory-worthy statements (POST /api/v1/memories/extract).
  3. Periodic chat insight extraction — a scheduled job sweeps workspace chat for operator preferences, guidance, and decisions. Automatically extracted memories are confidence-capped below manual ones so they never outrank what you stated explicitly.

How Memories Are Used

Each memory carries an importance (1–10) and a confidence (0–1). When an agent runs, a budgeted context block is assembled from the memories most relevant to the current query — ranked by embedding similarity with importance as the tiebreaker — and injected into the prompt. Workspace memories above a confidence threshold are similarly rendered into the workspace runtime prompt.

Memories can expire (expires_at) and can be archived instead of deleted when you want them out of prompts but kept for the record.

API Summary

EndpointPurpose
GET /api/v1/memoriesList, filterable by agent, user, and type
POST /api/v1/memories, PUT/DELETE /api/v1/memories/{id}Manage entries
POST /api/v1/memories/{id}/archiveArchive without deleting
POST /api/v1/memories/extractExtract memories from a conversation
GET /api/v1/memories/contextDebug: see the rendered context block agents receive