Use and Disuse: Intent-Structured Experience Consolidation for Memory and Learning in LLM Agents

Source facts

The preprint proposes Hippocam, a hierarchical memory and continual-learning architecture that organizes ongoing work as nested intents. Active context stays centered on the current intent, while completed intents are consolidated into task-relevant outcomes and state. A recursive prefix-consolidation mechanism abstracts long-unused experiences while preserving original interactions for later recovery. The authors describe capability evolution through experience without parameter updates.

Analysis

Hippocam makes memory maintenance selective and intent-conditioned rather than treating the full interaction history as equally reusable. The proposal is relevant to long-running agents, but the abstract does not establish how consolidation errors, stale knowledge, or privacy constraints behave over extended deployment.

Uncertainty/Doubt

• This is an arXiv preprint. The architecture's long-horizon behavior, consolidation quality, and claimed learning effects were not independently reproduced here.