Most AI memory discussions fixate on retrieval. How do you store knowledge so you can find it later? Vector databases, RAG pipelines, embedding similarity — all variations on the same question: given a query, what’s relevant?
But that’s only one kind of remembering. There’s another kind that gets almost no attention in AI agent design: remembering to remember. Psychologists call it prospective memory — the ability to recall an intention when a future context arises, not when you go looking for it.
The difference matters. “What’s Terry’s mortgage payment?” is retrospective — you search for it. “Next time we’re triaging email, flag any Grammarly renewal nudges” is prospective — nothing to search for yet. The trigger hasn’t happened. You need to recognise the context when it arrives and surface the intention at that moment.
The gap
I run Claude Code as my primary interface for… most things. It has a memory system with several layers:
- Rules (CLAUDE.md) — always loaded, permanent, unconditional
- Gotchas (MEMORY.md) — always loaded, high-frequency patterns to avoid
- Skills — loaded on demand, permanent capabilities
- Conversation — ephemeral, gone after the session
None of these handle “when X happens, do Y, then forget about it.” Rules are permanent. Conversation is too transient. Skills are for capabilities, not reminders. And MEMORY.md is for patterns, not one-off triggers.
So I added a fifth layer: a simple markdown file of context-triggered reminders.
- WHEN: email triage and seeing Grammarly emails → THEN: do NOT renew, lapsing Mar 28
- WHEN: working in ~/code/sopor → THEN: run migration for historical data
- WHEN: solving a client problem → THEN: ask "can I test this instead of arguing for it?"The format is deliberately simple. WHEN: <trigger context> → THEN: <action>. Each entry gets checked when entering a matching context — email triage, a project directory, a type of thinking. After it fires and the action is taken, delete it.
Why this works better than it should
Three properties make this disproportionately useful:
It bridges sessions. The subscription audit happened today. The email triage happens Thursday. Different sessions, different context windows. Without the prospective entry, Thursday-me has no idea that Tuesday-me decided to let Grammarly lapse. The decision falls through the crack between conversations.
It’s context-shaped, not time-shaped. A calendar reminder fires at 3pm whether or not you’re doing email. A prospective reminder fires when you’re actually in the right context to act on it. “Flag Surfshark renewal” is useless during a coding session. It’s perfectly timed during email triage.
It self-destructs. This is the key design choice. Prospective entries are temporary by definition. They exist to bridge a gap between a decision and its execution context, then disappear. No accumulation, no decay management, no staleness reviews. Fire once, delete, done.
The pattern underneath
Every personal productivity system eventually reinvents prospective memory. GTD’s “waiting for” list. Trigger-based reminders. “If I’m at the store, buy milk.” The insight isn’t that this is novel — it’s that AI agent architectures systematically neglect it.
The industry conversation about AI memory is stuck on “how do we remember more?” The more interesting question is: how do we remember at the right moment and then stop remembering?