Most discussions about AI memory are about how to store more: longer context windows, bigger vector stores, more history. For a personal assistant, the harder and more interesting question is the opposite one: what should it not keep, and what should it let fade?
This is a design note about the trade-offs. We're not presenting benchmarks, just the reasoning we find useful when thinking about personal memory.
Why "remember everything" is a trap
Retaining every message forever sounds safe. In practice it creates problems:
- Noise grows with volume. The more you store at equal weight, the more irrelevant material competes with what matters at retrieval time.
- Stale facts linger. Old addresses, cancelled plans and superseded preferences keep matching queries.
- Sensitive data accumulates. Everything retained is something that has to be protected, exported and deletable.
A simple taxonomy of what to keep
It helps to sort incoming information by how it will be used later:
| Type | Example | Handling |
|---|---|---|
| Commitments | "Remind me to send the invoice Friday" | Keep until done, then archive |
| Durable facts | Someone's birthday, a preference | Keep, but allow updates to supersede |
| Reference material | A saved article, a receipt | Keep retrievable, low priority in recall |
| Ephemeral chatter | "ok thanks", small talk | Don't store as memory |
| Context for now | What we were just discussing | Short-term only, expire quickly |
The short-term vs long-term split is the most important line. We cover it in more detail in how AI assistants remember across short-term and long-term memory.
Supersession beats deletion
Facts change. Rather than only appending, a memory system should be able to say "this replaces that". If the user says their dentist appointment moved, the old time shouldn't just sit alongside the new one. A pragmatic approach is to keep the old record but mark it superseded, so it stops competing in retrieval while remaining auditable.
Let importance be earned
Instead of guessing importance at write time, signals can accumulate:
- Was it mentioned again?
- Was it retrieved and used?
- Did the user correct or confirm it?
- Is it attached to a date or a person that recurs?
Things that never get touched can decay in priority. Things that keep coming up can be promoted.
Forgetting must be user-controlled
For personal data, "the system decided to forget" is not enough. Users should be able to see what is stored, edit it, and delete it for real. Two practical guides on the user side: how to delete your data from AI services and an AI data portability checklist.
Entity extraction as a filter
Structured extraction (people, places, dates) doubles as a relevance filter: a message with a date and a person is probably worth keeping; a greeting isn't. We described that side in how AI extracts people, places and dates from notes.
Takeaway
Good personal memory is curated, not hoarded: keep commitments and durable facts, expire context quickly, let updates supersede old values, and always leave the user in control. If you're building in this space, design the forgetting alongside the remembering.
We're applying these ideas in Brinn, a personal AI for reminders, lists and memory across WhatsApp, email, web and desktop. How do you handle forgetting in your own systems? We'd love to hear in the comments.
Top comments (0)