BMAM: Brain-inspired Multi-Agent Memory Framework

TL;DR

This paper tackles soul erosion, the degradation of temporal coherence, semantic consistency, and identity in long-horizon AI agents, by introducing BMAM, a brain-inspired multi-agent memory framework. BMAM decomposes memory into specialized subsystems (episodic, semantic, salience) coordinated by a central controller, and employs a StoryArc timeline for explicit temporal indexing along with a hybrid retrieval scheme that fuses lexical, semantic, relational, and temporal signals, enabling temporally grounded reasoning. Empirical evaluation on LoCoMo and LongMemEval shows BMAM achieves 78.45% accuracy on LoCoMo and 67.60% on LongMemEval, with ablations underscoring the critical role of the hippocampus-inspired episodic memory for temporal reasoning; error analysis highlights temporal confusion and entity ambiguity as primary challenges. The work presents a principled architectural pattern for persistent memory in AI agents and points to future directions in cross-modal memory, embodied agents, and adaptive component activation to further strengthen long-horizon memory and personalization.

Abstract

Language-model-based agents operating over extended interaction horizons face persistent challenges in preserving temporally grounded information and maintaining behavioral consistency across sessions, a failure mode we term soul erosion. We present BMAM (Brain-inspired Multi-Agent Memory), a general-purpose memory architecture that models agent memory as a set of functionally specialized subsystems rather than a single unstructured store. Inspired by cognitive memory systems, BMAM decomposes memory into episodic, semantic, salience-aware, and control-oriented components that operate at complementary time scales. To support long-horizon reasoning, BMAM organizes episodic memories along explicit timelines and retrieves evidence by fusing multiple complementary signals. Experiments on the LoCoMo benchmark show that BMAM achieves 78.45 percent accuracy under the standard long-horizon evaluation setting, and ablation analyses confirm that the hippocampus-inspired episodic memory subsystem plays a critical role in temporal reasoning.

Figures (13)

• Figure 1: Soul erosion types and BMAM countermeasures. Each erosion mechanism requires a specialized defense: temporal erosion is addressed by StoryArc timeline indexing, semantic erosion by hippocampus-to-temporal-lobe consolidation, and identity erosion by amygdala salience tagging.
• Figure 2: BMAM architecture overview. A central coordinator orchestrates multiple functionally specialized memory subsystems sharing a unified memory substrate with episodic timelines, a knowledge graph, and vector-based storage.
• Figure 3: LoCoMo benchmark comparison. BMAM achieves 78.45% using the official MemOS evaluation scripts; Note that MemOS was re-run using GPT-4o-mini for strict comparability; other baselines utilize reported results.
• Figure 4: Brain-region ablation on LoCoMo. Hippocampus removal causes a 24.62% drop, validating episodic memory as the critical backbone. Varied effects for other components reflect tight coupling (see text).
• Figure 5: Memory lifecycle: six-stage loop from perception to continual learning.