Memory and self-learning

Memory is markdown files in the agent's workspace, which means you can read and correct it directly — in the dashboard's Memory tab, the Files tab, or a text editor.

• MEMORY.md — durable facts the agent has learned and should keep.
• memories/YYYY-MM-DD.md — a daily log. Today's and yesterday's are loaded into every conversation.

Self-learning (memory.self_learning, on by default) is part of the Cortex engine, so it runs only for an agent Cortex is on for, and Cortex is off until you turn it on (Cortex and Pulse). While it runs it reads recent conversations, distils what is worth keeping, and writes it into the daily log and MEMORY.md. A weekly pass consolidates. self_learning_interval (default 3600 seconds) sets the pace.

The same background pass (so, again, only with Cortex on for the agent) also keeps an Activity record in the daily log — what the agent did, as opposed to what it learned: each scheduled run, heartbeat finding, request from another agent, and Cortex phase, one line each with the time and a short excerpt of the request and the reply. It is written deterministically (no model call, no credits) on every cycle, so it is there even on days when the distillation itself is skipped, and it is bounded — one block per day, capped in length — so a busy day cannot crowd out the rest of the prompt. Ordinary conversations with people are not copied into it; those go through distillation only. Ask the agent what it did yesterday and it reads this block; the dashboard Overview → Daily highlights shows the fleet-wide version of the same day.

Turning on memory.rag also indexes memory for meaning-based search, so the agent can find something it learned months ago without it being in the prompt.

Per-customer memory (memory.per_customer) is the multi-user variant — see Multi-user: two different features.

If an agent has learned something wrong, editing MEMORY.md is the fix, and it takes effect on the next reply.