Prospective Memory: Remembering the Future, Not Just the Past

Most memory systems are good at recalling the past. Prospective memory lets an agent hold an intention that surfaces at the right moment.

Ask an agent what last week's meeting was about, and it will usually answer just fine. But tell it "remind me about this again next Wednesday," and most memory systems go quiet — because they were built to look backward, and only backward.

Human memory doesn't work that way. We remember not only what has happened, but what hasn't happened yet and needs to: the letter to mail before leaving, the quarterly review due at month's end, the unanswered request to chase in three days. Psychologists call the second kind prospective memory — holding a future intention and letting it rise to mind at the right time. It's exactly the half that Corsoul sets out to complete.

The idea: a prospective intention

At the heart of prospective memory is a prospective intention: remembering that at a certain moment, you should do — or recall — something.

What sets it apart from ordinary memory is that it doesn't describe something that already happened. It describes a commitment that hasn't fired yet. A prospective memory binds two things together in advance: a trigger (a point in time, or a recurring cadence) and the content of the intention. Until the trigger arrives it simply waits, quietly, taking up no attention and demanding nothing. When the moment comes, it becomes eligible to surface and to remind the agent: this is the thing, and now is the time.

The difference matters because backward-looking recall alone leaves a gap. An agent can know everything about what it has done and still drop the ball on what it promised to do — not from forgetting the fact, but from having nowhere to park a future obligation until its moment arrives. Prospective memory is that place.

For a long-running agent, this is the other side of a continuous life history. Past memory tells it who it is and what it has learned. Prospective memory tells it what it still owes and where it's headed next.

Three plain actions: register, check what's due, close out

Corsoul distills prospective memory into three plain actions:

• Register an intention — store it. It can be one-off ("chase this in three days") or recurring ("review every Monday").
• Check what's due — ask "what's due right now?" and pull back the intentions that are ready to surface at this moment.
• Close it out — mark it done, or reschedule it to its next occurrence.

A typical loop looks like this. The agent promises something mid-conversation, so it registers the intention. On each beat afterward, the agent checks whether anything has come due. Once it's handled, it closes the intention out — and a recurring one automatically lines up its next turn.

A passive invariant: memory never acts for you

Here is a line Corsoul deliberately holds: checking what's due is just a look. You can ask as often as you like and it won't change any state — querying is not claiming, and it is certainly not executing.

More importantly, this layer only stores intentions; it never carries out an external action on its own. It won't send your email, won't call an API for you, won't fire anything in the background. It does exactly one thing: at the right moment, it hands the right intention back to the agent.

Whether to act, when to act, and how to act all stay on the agent's side. This isn't an unfinished feature — it's a hard invariant by design: the cadence of the loop and the actual action belong to the agent alone. Precisely because the memory layer holds no executor and never pushes, it can be plugged safely into any runtime, and it can serve as a clean namespace in a multi-tenant setting. Memory remembers; action stays where action belongs.

Free on every tier — and where it fits

Worth stating plainly: prospective memory is available on every tier of Corsoul. Reminders are the most basic form of memory, not a paid brain feature gated behind the licensed engine — even on the free, objective memory layer, all three actions — registering an intention, checking what's due, and closing it out — are fully open.

It fits a few rhythms especially well:

• Cross-session to-dos — something promised in this turn but due in the next (or the next day) won't evaporate when the conversation ends.
• Recurring checks — a daily sweep, a weekly review, a monthly close: one recurring intention keeps the cadence.
• The pacing of long-running work — for a task that spans several days, you can plant the checkpoints worth revisiting so the agent recalls them at the right time on its own, instead of hoping the thread stays alive in context.

Memory shouldn't be only an archive of the past. For an agent meant to work alongside you over the long haul, it has to remember the future too. That's the local-first Corsoul approach: start free, let it hold both what happened and what's still to come — and when you need deeper capabilities like associative learning and emergent patterns, light up the paid personality engine.

Start free. Memory becomes experience. Experience becomes a self.