Start free
Docs / AI Agent / Memory system
Explanation

Memory system

A good assistant remembers you — you don't have to explain from scratch every time. SemAgent's memory has three layers (semantic · episodic · procedural), plus ambient proactivity and skill distillation on top. All of it is consent-based and viewable/deletable from the GDPR panel.

🔒

Consent gates: memory features don't run without env + tenant + user consent. Each layer can be toggled independently; everything learned is visible and deletable in one click.

1. Semantic profile

Your preferences and working style are inferred from conversations: the tone you like, the tasks you do regularly, the constraints you care about. This inference runs in the background ("dreaming"); a distilled profile is stored, not raw chat. The agent uses it as context on the next task — but the profile is a frozen snapshot; you view, edit or delete it.

2. Episodic memory

  • Daily/weekly summaries are generated automatically — "what got done this week" at a glance.
  • The agent recalls the past with recall_history: "continue last week's report task" works.
  • The last few days' summary is injected into context, preserving continuity without wasting tokens.

3. Procedural rules

When you make the same correction twice ("always attach invoices as PDF"), the agent promotes it to a temporary candidate rule; once re-confirmed it becomes an active rule and is applied on its own from then on. Rules are per-user and managed from the panel (view / pause / delete). This is the "observe → correct → repeat → promote to rule" loop.

Ambient proactivity

Memory is the basis for making the agent proactive: notice → suggest → approve → act. For example the personal morning briefing feeds from your memory (follow-ups, commitments). Proactivity level is adjustable (suggest-only → act+notify) and the safety hard-floor holds — even when proactive, a risky action still requires approval. More: Scheduling & briefings.

Skill distillation

From successful tasks the agent distills a playbook: when a similar job comes up again it remembers the steps and gets sharper each time. Skills are visible in the panel.

GDPR & control

  • Everything learned (profile, summaries, rules) is visible in the GDPR panel.
  • One-click deletion; separate consent/toggle per layer.
  • Memory + skill cost is transparent — you see it in the usage panel.
Was this page helpful? Send feedback