LeanMem: Simple and Efficient Long-Term Memory for LLM Agents

LLM-based agents often struggle with long-term memory due to the tension between high token costs and the loss of fine-grained information. Existing systems typically use a uniform pipeline to process all dialogue, which is either too expensive or too reductive. This paper asks: can we improve memory efficiency and fidelity by treating different types of conversational information—such as stable preferences, evolving events, and dense records—with distinct, specialized storage and retrieval strategies?

LeanMem introduces a three-stage framework designed to handle information based on its inherent properties:

1. Controlled Memory Writing: Instead of uniform processing, the system filters out low-value turns and uses an LLM scheduler to classify segments. Stable attributes are stored as compact profiles, state-changing events are stored with temporal anchors, and detail-intensive content is stored as records with source pointers.
2. Selective Memory Evolution: Rather than constantly updating the entire memory bank, LeanMem performs deferred, localized updates only on event memory, amortizing costs while keeping state tracking accurate.
3. Adaptive Evidence Composition: At inference time, the system analyzes the query to determine the specific evidence demand (e.g., needing a stable fact versus a chronological history). It then allocates a retrieval budget dynamically across the different memory types, assembling only the necessary evidence.

LeanMem was evaluated on the LoCoMo and LongMemEval-S benchmarks using both GPT-4.1-mini and Qwen3-8B. The results show that LeanMem consistently outperforms existing memory-based baselines, achieving accuracy improvements of up to 15.1 points. Simultaneously, it significantly reduces construction costs, inference token usage, and latency, demonstrating that a heterogeneous approach to memory management is more effective than a one-size-fits-all pipeline.