AI Memory (Self-Updating)

What AI Memory is

AI Memory lets the assistant remember things about you across conversations — preferences, facts, recurring context — so you don't have to repeat yourself. It's self-updating: memories are formed from your interactions, kept fresh, and decay over time if they stop being relevant, so stale facts fade instead of lingering forever.

How it works

• Memory items capture a fact or preference worth remembering.
• Each item is embedded so it can be retrieved by semantic relevance, not just exact wording.
• Items track their source (where the memory came from) and a relevance/decay signal, so the system can prefer recent, reinforced memories and let unused ones fade.

Memory is user-scoped — your memories are yours and are used to personalize your own sessions.

Retrieval during a conversation

When relevant, the assistant retrieves memory items by semantic similarity to the current conversation and uses them as context. Because retrieval is embedding-based, a memory like "prefers metric units" can surface even when you phrase the request differently.

Decay and freshness

Memories aren't permanent by default. A memory that keeps being relevant is reinforced; one that's never used again decays and eventually stops influencing answers. This keeps the assistant's picture of you current rather than cluttered with one-off details.

Source tracking

Every memory records where it came from. This makes memory auditable — you can understand why the assistant "knows" something rather than it appearing from nowhere.

Privacy

• Memory is personal data and is covered by ToRun's data-subject tooling: you can request export or erasure of your data, and content is hard-deleted on erasure.
• Multilingual matching is language-agnostic, so a memory stored in one language can still match a query phrased in another.

Tips

• State durable preferences explicitly ("always answer in Turkish," "I work in finance") so they're captured cleanly.
• If something changes, just tell the assistant the new fact — reinforcement and decay handle the transition over time.
• Use export/erasure controls if you want to review or clear what's been remembered.