Agentic memory is becoming essential for long‑horizon AI agents, yet most existing systems let autoregressive large language models (LLMs) directly control how memories are organized, retrieved, and used, putting expensive generation on the critical path of memory operations.
We introduce Jev-Mem, a new agentic memory architecture inspired by the System‑One/System‑Two cognition model. System One handles fast, lightweight decisions, while System Two performs slower, deliberative reasoning. Jev‑Mem maps this division of labor onto three planes:
- System‑One control plane: during memory construction it determines memory typing and multi‑relational organization; during retrieval it dynamically performs query routing, retrieval‑budget allocation, graph traversal, candidate scoring, and adaptive stopping.
- Structured multi‑relational memory plane: stores entities and their multiple relations in a graph, enabling efficient neighbor queries.
- System‑Two reasoning plane: invoked only for complex reasoning or answer synthesis, avoiding unnecessary LLM calls.
This design improves both memory effectiveness and system efficiency. On the LoCoMo benchmark Jev‑Mem achieves an overall LLM‑as‑a‑Judge score of $0.777$, a $11.0\%$ relative improvement over the strongest baseline, reduces memory construction time to $158\,\text{s}$ (a $6.6\times$ speedup), and lowers average query latency to $0.93\,\text{s}$ (a $36.7\%$ reduction).
Review: By cleanly separating System‑One and System‑Two responsibilities, Jev‑Mem removes high‑cost model inference from routine memory operations, delivering substantial speed and performance gains and offering a practical memory framework for large‑scale, long‑term AI agents.