Reconstructive Episodic Memory: A Functional Approach to Memory for LLM Agents

We present Reconstructive Episodic Memory (REM) — a functional approach to memory for large language model (LLM) agents. Instead of retrieving stored data from a database or vector index, REM encodes each memory episode into a small neural function that can deterministically reconstruct the original information from a semantic key. This paradigm shifts memory from a passive storage model to an active, reconstructive process, closer to how biological memory operates. We describe the theoretical foundations, implementation details, and experimental evaluation of REM, demonstrating byte-exact recall, linear scalability, and zero-knowledge-like behavior. The results suggest that functional memory can become a fundamental component of future LLM architectures and autonomous cognitive systems.