Memories

4 min read

Memories are the persistent context of agents: policies, product information, decisions, internal procedures, or customer data. They are managed as documents and can be assigned to one or more agents.

Memory scope

Type	Use
Core	The agent's identity and canonical guidance. The system creates these and they cannot be unassigned.
Own	Context specific to a single agent.
Shared	Context reusable by several agents in the organization.

A shared memory is preferable when the same information must remain consistent across agents, such as commercial policies or the service catalogue.

Short- and long-term memory

Each memory has a time scope that determines how it takes part in the agent's work:

• Short term: loaded when the agent starts and forms part of its working context. Use it for instructions, criteria, and information it must always consider.
• Long term: kept as knowledge for reference and is not included in the initial context. It is appropriate for extensive, historical, or case-specific material.

Write good descriptions

For long-term memories, the description helps identify their content and when they can be useful. State both, for example: “Current prices, discounts, and terms for Spain”.

An agent that maintains its context

You do not have to create and update every memory manually. If you enable the Memories and skills capability, the agent can create documents when it detects information worth keeping, update knowledge that has changed, and archive what no longer helps.

This lets context evolve with the agent's real work: a new policy, a decision made in a conversation, or an improved procedure do not have to wait for manual maintenance.

With the same capability, the agent also tidies its memory when you ask it to, or when what it loads on every step starts to weigh too much: it merges repeated memories, summarises long ones, moves to long-term the ones it does not always need, and archives the ones it no longer uses. From a shared memory it can only unassign itself: the other agents keep it unchanged. Core memories can only be summarised.

Weekly review

With the Memory review card turned on under Tools in the agent configuration, the agent reviews itself every Monday before dawn, in its own time zone: it is the same review as AI Doctor. On every plan it looks for personal information about a team member in a memory everyone can read, and for passwords or keys written in a memory. On Pro, also for repeated, outdated, overly long or undescribed memories, and undescribed skills.

It changes nothing on its own. Everything it proposes in a review reaches you as a single approval, one change per row: what moves, where to (that person's private notes, a new credential in the vault, a merged memory or the archive) and what the memory looks like afterwards. Tick the ones you want applied; the ones you untick are not proposed again. Memories it shares with other agents are left alone. The approval is decided by whoever can edit memories, even when the organization is set to "Automatic"; moving a password or key to the vault also requires being able to edit credentials. If a memory changes before you approve it, that change is not applied and the rest are.

The setting is off by default on every agent.

Define the framework, not every document

Keep core memories and access rules under human control. Let the agent manage operational knowledge within those limits. You can block or grant this capability from Behaviour and limits.

Organize and maintain

Use folders to group context by area, customer, or team. Search and filters let you find memories by type, folder, or assignment status; archived items temporarily retain removed memories so you can recover them.

Periodically review memories with no assigned agent and those that are outdated. Small, current, well-classified context is usually more useful than an extensive collection without maintenance. When an agent has accumulated many memories, run AI Doctor from its tab: it analyzes its memories together with the rest of its configuration and suggests what to review or adjust.

Changes and versions

Each edit creates a version that you can view or restore. Retention and the way versions are kept depend on the plan and are configured in Organization.

Memories and skills

Memories

Context: information, criteria, and decisions the agent needs to know.

Skills

Procedures: instructions and resources for carrying out a specific task.