TLDR: PISA is a novel AI agent memory system inspired by Piaget’s cognitive theory, designed to make AI agents more adaptive and capable of continuous learning. It uses a “Schema Engine” to organize memory based on tasks and dynamically updates this knowledge through assimilation, schema evolution, or creation. PISA also features a hybrid retrieval system combining symbolic reasoning with neural retrieval, significantly improving how AI agents access and use information. It outperforms existing systems on conversational and data analysis tasks, including a new benchmark called AggQA.
In the rapidly evolving world of Artificial Intelligence, Large Language Model (LLM) based agents are becoming increasingly sophisticated, capable of autonomously tackling complex objectives. A critical component enabling their intelligence is an effective memory system. However, existing AI memory solutions often struggle with adaptability across diverse tasks and tend to treat memory as a passive storage unit rather than an active, constructive process.
Addressing these limitations, researchers from Zhejiang University have introduced PISA: A Pragmatic Psych-Inspired Unified Memory System for Enhanced AI Agency. This innovative framework draws inspiration from Jean Piaget’s theory of cognitive development, which posits that knowledge is actively built through interaction with the environment.
The Core Idea: Memory as an Adaptive Process
PISA redefines AI agent memory by treating it as a dynamic, constructive, and adaptive process. Unlike systems that merely store information, PISA’s memory actively organizes and evolves, allowing AI agents to continuously learn and adapt to new situations. This is crucial for overcoming challenges like overwhelming context windows and interference from irrelevant historical data, which can hinder an LLM’s reasoning and decision-making.
PISA’s Architecture: A Schema-Driven Approach
At the heart of PISA is its uniquely designed Schema Engine. This engine organizes memory hierarchically into a Memory Pool, Buckets, and Schemas. Each schema is a structured unit of knowledge, defining high-level topics, specific elements, and nuanced descriptors (like ‘Attitude: like’ or ‘Scenes: Winter morning’). This structured yet flexible design ensures that memory is task-oriented from the outset.
The system comprises three main modules:
-
Initialization Module: This module proactively constructs PISA’s memory structure by analyzing the AI agent’s goals and task specifications. It sets up task-oriented memory ‘buckets’ (e.g., User Trait, User Event) and defines schema-level element-value structures, ensuring that memory is aligned with the agent’s objectives.
-
Adaptation Module: Inspired by Piaget’s concepts of assimilation and accommodation, this module enables continuous learning. When new experiences arrive, PISA dynamically decides whether to:
-
Assimilate: Integrate new information into an existing schema (e.g., adding new details to a ‘coffee’ schema).
-
Accommodate (Evolve): Modify an existing schema by adding new elements if current ones are insufficient (e.g., adding ‘Milk’ to a ‘Drink’ schema).
-
Accommodate (Create): Form an entirely new schema for novel concepts that cannot be explained by existing structures (e.g., creating a ‘Music’ schema).
This trimodal mechanism ensures coherent knowledge organization and flexible memory updates.
-
-
Retrieval Module: Designed as a ReAct AI agent, this module interprets user queries and retrieves relevant information. It uses an extensible toolkit library, including Retrieval-Augmented Generation (RAG) for factual queries, SQL for structured data reasoning, and a Calculator for aggregation tasks. This hybrid approach seamlessly integrates symbolic reasoning with neural retrieval, significantly boosting retrieval accuracy and efficiency.
Also Read:
- RGMem: A Physics-Inspired Framework for Evolving User Profiles in Language Agents
- AUGUSTUS: An AI Agent with Human-Like Multimodal Memory
Empirical Validation and a New Benchmark
The effectiveness of PISA was rigorously evaluated on existing benchmarks like LOCOMO (for long-term conversational memory) and a newly proposed benchmark called AggQA. AggQA is specifically designed for data analysis tasks in practical domains such as Medical and Finance, addressing a gap in current benchmarks that often lack task diversity and real-world relevance.
PISA consistently outperformed all baseline models across various evaluation metrics, demonstrating significant improvements in adaptability and long-term knowledge retention. The ablation studies further confirmed the critical roles of both the Initialization and Retrieval Modules in PISA’s superior performance.
This research marks a significant step towards building more intelligent and adaptive AI agents, offering a memory system that learns and evolves much like human cognition. For more in-depth details, you can refer to the full research paper: PISA: A Pragmatic Psych-Inspired Unified Memory System for Enhanced AI Agency.
