PISA: A Pragmatic Psych-Inspired Unified Memory System for Enhanced AI Agency

TL;DR

PISA addresses the limitations of task-agnostic AI memory by introducing a Piaget-inspired, unified memory system that treats memory as an active, constructive process. It combines a Schema Engine with tri-modal adaptation (assimilation, accommodation, creation) and a hybrid retrieval architecture that integrates symbolic reasoning and neural retrieval through a LangGraph-based framework. Empirical results on LOCOMO and AggQA demonstrate state-of-the-art adaptability and long-term knowledge retention, with ablations validating the importance of initialization and retrieval structure. The work advances practical AI agency by enabling task-oriented memory organization, continuous learning, and robust, multi-faceted data reasoning, while highlighting avenues for broader benchmarking and priors-based improvements.

Abstract

Memory systems are fundamental to AI agents, yet existing work often lacks adaptability to diverse tasks and overlooks the constructive and task-oriented role of AI agent memory. Drawing from Piaget's theory of cognitive development, we propose PISA, a pragmatic, psych-inspired unified memory system that addresses these limitations by treating memory as a constructive and adaptive process. To enable continuous learning and adaptability, PISA introduces a trimodal adaptation mechanism (i.e., schema updation, schema evolution, and schema creation) that preserves coherent organization while supporting flexible memory updates. Building on these schema-grounded structures, we further design a hybrid memory access architecture that seamlessly integrates symbolic reasoning with neural retrieval, significantly improving retrieval accuracy and efficiency. Our empirical evaluation, conducted on the existing LOCOMO benchmark and our newly proposed AggQA benchmark for data analysis tasks, confirms that PISA sets a new state-of-the-art by significantly enhancing adaptability and long-term knowledge retention.

Figures (13)

• Figure 1:
• Figure 4: Overall architecture of PISA. The system is built upon a uniquely designed Schema Engine and consists of three core modules: (1) the Initialization Module, which leverages agent goals and prior knowledge to initialize memory structures; (2) the Adaptation Module, which dynamically processes new experiences through schema assimilation, evolution, and creation, guided by conflict analysis; and (3) the Retrieval Module, which interprets queries and retrieves relevant memories using task-aware retrieval strategies and external tool calls. Together, these components enable a flexible, extensible, and task-oriented memory system for LLM-based AI agents.
• Figure 5: Schema Engine of PISA. The memory system is organized hierarchically into Memory Pool, Buckets, and the Schema. Each schema encodes structured knowledge through Meta, Element, Element-Value, and a unique experience ID, enabling flexible, extensible, and task-oriented memory organization.
• Figure 6: Retrieval Module of PISA. It supports diverse query handling through Regional Fact Retrieval, Multi-Fragment Reasoning, and Aggregation, leveraging an extendable toolkit library (e.g., RAG, SQL, and Caculator) to utilize both unstructured and structured memory.
• Figure 7: Analysis of the impact of $\theta_{\text{meta}}$ on agent performance and internal behavior. C1 to C4 represent Single-Hop, Multi-Hop, Temporal, and Open-Domain tasks, respectively.