Overview
Liu et al. (UNC-Chapel Hill, UC Berkeley, UC Santa Cruz; arXiv:2601.02553, Jan 2026) — Efficient memory framework for LLM agents based on semantic lossless compression. Achieves 26.4% F1 improvement over baselines while reducing inference token consumption by 30x.
Three-Stage Pipeline
1. Semantic Structured Compression
Entropy-aware filtering preserves high-utility information while discarding redundant content. Uses an information score combining entity novelty and semantic divergence:
H(Wt) = α · |Enew|/|Wt| + (1-α) · (1 - cos(E(Wt), E(Hprev)))
Windows below threshold are excluded from memory construction entirely.
2. Recursive Memory Consolidation
Asynchronous process inspired by biological memory consolidation (Complementary Learning Systems theory). Related memory units are recursively integrated into higher-level abstract representations, reducing semantic redundancy.
3. Adaptive Query-Aware Retrieval
Dynamically adjusts retrieval scope based on estimated query complexity. Lightweight symbolic and semantic constraints prune irrelevant candidates.
Results (LoCoMo Benchmark)
| System | F1 Score | Token Cost |
|---|---|---|
| SimpleMem | ~48 | ~550 tokens |
| LoCoMo (Full) | ~44 | ~16,500 tokens |
| Mem0 | ~22 | ~550 tokens |
| MemGPT | ~28 | ~6,000 tokens |
SimpleMem occupies the ideal high-accuracy/low-cost position.
Relevance to AP
Directly applicable to the knowledge vault’s context management challenges. The three-stage pipeline (compress → consolidate → retrieve) maps to how research materials should flow through the integration pipeline.
Related
- deepseek-ocr-context-compression — Complementary optical approach to context compression
- compact-guide-to-rag — RAG architecture that could incorporate SimpleMem’s retrieval
- deepseek-conditional-memory — Related memory mechanism research