Introduction
This paper introduces "preemptive strategies" to solve the "chicken-and-egg" problem of prospective memory in artificial agents: how can an agent remember to perform a goal if it has already forgotten the goal exists? By using periodic or context-based triggers to proactively rehearse or reactively retrieve goals from long-term memory, agents can successfully navigate complex environments without losing track of their long-term objectives.
In cognitive architectures like Soar and ACT-R, forgetting is a functional feature used to manage the "utility problem," preventing an overflow of information from slowing down decision-making. However, forgetting becomes a liability for prospective memory—the ability to remember to perform an action in the future (e.g., "buy milk on the way home").
The challenge lies in a causal loop: to recognize an opportunity to act (such as seeing a grocery store), the relevant goal must reside in the agent's active working memory. If a goal has been suspended for a long duration, it likely decayed and moved to long-term memory. Without the goal in working memory, the agent fails to recognize the opportunity, and the goal remains "forgotten." This research explores "preemptive strategies" to break this loop without altering the fundamental memory handling of cognitive architectures.
Methods
The authors propose two classes of preemptive strategies that rely on goal-independent triggers, implementable via standard if-then rules:
- Proactive Maintenance: The agent "rehearses" uncompleted goals at specific triggers to boost their base-level activation, preventing them from decaying out of working memory.
- Reactive Maintenance: The agent allows goals to decay but uses triggers to deliberately retrieve them from long-term memory back into working memory.
Two types of triggers were evaluated:
- Timing-based: A periodic "alarm" (e.g., every 400 motor steps) prompting a goal check.
- Domain-based (Context-switch): A trigger tied to environmental changes, such as entering a new room, which often predicts new opportunities.
Experimental Setup
The strategies were tested using a simulated SuperDroid robot in a multi-room indoor environment. The robot followed a fixed patrol route with the task of picking up and delivering objects to specific rooms. Deliveries were pre-loaded into long-term memory, requiring the robot to recognize objects while patrolling. The experiments varied the memory decay rate and the robot's speed to test the robustness of the strategies.
Results and Discussion
The study revealed several critical insights into memory management:
- The Scaling Wall: There is a mathematical limit to goal management. If too many goals are retrieved at once, the first may decay before the agent finishes processing the last.
- Predictiveness vs. Frequency: Timing triggers are reliable but inefficient. Domain triggers (like room entry) are computationally "cheaper" because they fire less often, but they require high predictiveness to be effective.
- Robustness of Retrieval: Reactive retrieval is generally more robust than proactive rehearsal. Rehearsal requires precise calibration of decay rates; retrieval is more forgiving as the agent can simply retrieve the goal whenever a trigger occurs.
Key results from the robot simulation showed that timing triggers maintained 100% success up to a decay rate of 0.38. While the room-entry trigger was more efficient—with the timing trigger requiring 1.5x more retrievals at a decay rate of 0.54—it was more sensitive to the robot's speed.
Conclusion
While effective, these strategies currently involve retrieving all pending goals, which may not scale to thousands of tasks. Future work should investigate "metamemory"—an internal sense of memory decay—to trigger maintenance dynamically, and more complex, multi-cue triggers to improve retrieval efficiency in real-world scenarios.
Summary updated