Memory & learning
The difference between an automation and an employee is that an employee gets better. Syntrum agents remember what they learn, pick up new skills, and turn your coaching into lasting behavior — so the same mistake isn't made twice and the same question isn't asked again. This page explains what an agent remembers, how learning happens, and the workspace toggles that govern it.
Memory is the thread that connects the agent loop: the Report phase of one cycle feeds the Understand phase of the next. Learning is what makes that thread accumulate value over time.
What an agent remembers#
An agent's memory comes in a few flavors, each shaping future decisions:
| Kind of memory | What it captures | Example |
|---|---|---|
| What it learned | Patterns and facts the agent figured out on its own | "This contact prefers email over calls, usually replies Tue–Thu." |
| How you trained it | Explicit rules and behavior you taught it | "Always mention the ROI calculator in a first email." |
| Boundaries | Limits it must respect | "Never send a discount without approval." |
| Decisions | Your approve/reject patterns, so it can anticipate you | "Messaging a known contact was approved — do it directly next time." |
Memories carry confidence (from "just noticed once" to "always true") and scope (a single person, a company, or a situation), so the agent knows how much weight to give them. How long different memories last is shaped by an agent's memory policy, set in the guided builder — for example, "mappings weekly; decisions 90 days; incident patterns until superseded." See Hire your first agent.
How learning happens#
Agents pick up knowledge through three channels.
1. From their own observations#
As an agent runs its loop, it notices patterns — who prefers what, when replies tend to come, which approaches work. It records these as observations, and uses them the next time it faces a similar situation. This is the quiet, continuous kind of learning that happens without you doing anything.
2. From your decisions#
Every time you approve or reject an action in the Approvals queue, you're teaching the agent where your trust boundaries are. Approve a certain kind of action enough and the agent learns it's welcome; reject it and the agent learns to hold back. A vivid case: the first time an agent messages a new contact it asks — but once you approve, that contact is known, and it won't ask again.
3. From your chat guidance#
You can coach an agent directly in the inbox. When you give substantive guidance — "follow up after two days instead of one," "use a warmer tone with enterprise leads" — that instruction can become durable memory, not just a one-off reply. The agent carries it forward into future work.
The practical upshot: coaching a Syntrum agent feels like coaching a person. Tell it once, and — with learning enabled — it remembers.
Learning skills for tools#
Beyond facts and preferences, agents can teach themselves how to use the tools you connect. When a new integration is attached, an agent can work out how to operate it effectively rather than needing you to script each capability. This is how a freshly connected tool becomes genuinely useful to an agent quickly.
The learning toggles#
Because learning shapes behavior, it's governed at the workspace level. Open Safety (the AI Governance page) and select the Learning Behavior tab. Four switches control it:
| Toggle | What it does |
|---|---|
| Learn from my decisions | Agents remember your approve/reject patterns. |
| Learn tool skills automatically | Agents teach themselves how to use connected tools. |
| Remember chat guidance | Substantive chat replies become durable memory. |
| Adapt its own schedule | Agents may change their own timing — and every such change is approval-gated. |
These apply to every agent in the workspace. Turn them off and agents stop accumulating that kind of memory; turn them on and your workforce compounds what it knows.
Changing learning behavior is an admin/owner setting, since it affects every agent. If the controls are read-only for you, ask a workspace owner or admin. See Team & roles.
Memory and safety together#
Learning never overrides your guardrails. An agent that has "learned" it can message people freely still can't send a payment or a password-reset link if the workspace blocks those — memory makes an agent smarter, but the guardrails remain the floor. And because every action is recorded, you can always trace why an agent did something back to a specific memory or decision in the Audit log.
What's next#
- Guardrails & sensitive actions — where the learning toggles live and how the floor works.
- Chatting with agents — coach an agent and turn guidance into memory.
- The agent loop — see how Report feeds back into Understand.