Are We Ready for Agent-Native Memory?

Most current evaluations treat agent memory as a black box, measuring only end-to-end task success. This paper takes a data management perspective and decomposes agent memory into four core modules:

•⁠ ⁠Representation & Storage

•⁠ ⁠Extraction

•⁠ ⁠Retrieval & Routing

•⁠ ⁠Maintenance

Through a systematic evaluation of 12 representative memory systems across diverse workloads, the authors find that no single architecture dominates.

Effectiveness depends heavily on how well the memory design aligns with the specific bottlenecks of the workload (e.g., long-horizon stability, frequent knowledge updates, or procedural execution).

Key insights include:

•⁠ ⁠Structured and relational approaches often perform better for long-term consistency and updates.

•⁠ ⁠Localized maintenance tends to be more cost-efficient than global reorganization.

•⁠ ⁠Many existing systems still struggle to meet the full lifecycle demands of autonomous, long-running agents.

This Ledger Entry expands how readers think about infrastructure for agentic AI by showing that effective memory systems for long-horizon agents require careful alignment between modular design choices (representation, extraction, retrieval, and maintenance) and workload-specific bottlenecks — moving beyond generic RAG or long-context approaches toward truly agent-native architectures.