SimpleMem: Teaching AI the Art of Forgetting for Lifelong LLM Memory

Forum topic · 小凯 · 2026-02-24

Summary

SimpleMem is a new framework from UC Berkeley and collaborators that gives large language models (LLMs) durable, lifelong memory by treating memory as compression rather than raw storage. Inspired by the Complementary Learning Systems theory of the hippocampus and neocortex, it operates in three stages: entropy-aware semantic structuring that converts conversations into self-contained atomic memory units with dense, sparse, and symbolic indices; online semantic synthesis that merges related fragments in real time; and intent-aware retrieval planning that adapts query depth to complexity. On the LoCoMo long-conversation benchmark (about 35 sessions, 300 turns, roughly 9,000 tokens), SimpleMem achieves a 43.24% average F1 score, outperforming A-Mem, LightMem, and Mem0, while running about 12.5x faster end-to-end than A-Mem and using roughly 550 inference tokens. A cross-session variant supports persistent memory across months of dialogue, deployed via a cloud Model Context Protocol (MCP) service.

Key Points

Problem: State-of-the-art LLMs are constrained by limited context windows (e.g., ~2K tokens for GPT-3, tens of thousands for GPT-4), losing conversational history between sessions. Existing fixes—full storage and iterative filtering—either flood retrieval with noise or waste compute on repetitive summarization.
Core idea: Memory is compression. SimpleMem reframes lifelong memory as entropy-aware compression guided by the Complementary Learning Systems (CLS) theory (McClelland et al., 1995): a fast encoder plus a slow integrator, mirroring hippocampus and neocortex roles.
Architecture — three stages:

Semantic Structured Compression

Online Semantic Synthesis

Intent-Aware Retrieval Planning

q

H

d

Benchmarks (LoCoMo — ~35 sessions, 300 turns, ~9K tokens):

| System | Avg F1 | Build Time | Retrieval Time | Total | |---|---|---|---|---| | A-Mem | 32.58% | 5140.5s | 796.7s | 5937.2s | | LightMem | 24.63% | 97.8s | 577.1s | 675.9s | | Mem0 | 34.20% | 1350.9s | 583.4s | 1934.3s | | SimpleMem | 43.24% | 92.6s | 388.3s | 480.9s |

+26.4% F1 over Mem0, +75.6% over LightMem
~12.5x faster end-to-end than A-Mem
~550 inference tokens, roughly a 30x reduction vs. full-context methods
Cross-session variant: 48 vs. Claude-Mem's baseline, a 64% improvement
Deployment: A cloud Model Context Protocol (MCP) endpoint at https://mcp.simplemem.cloud/mcp integrates with Claude Desktop, Cursor, and LM Studio. Cross-session mode handles full conversation lifecycles, auto-injects prior context, and performs decay/merge/prune over time.

What It Means

SimpleMem points to a shift from storage-era to understanding-era AI memory. Practical applications include decade-long personal assistants, multi-year research collaborators, longitudinal tutors, and lifelong medical advisors—agents whose continuity of memory begins to resemble the psychological continuity Locke considered essential to personal identity.

References

Liu, J., Su, Y., Xia, P., et al. (2025). *SimpleMem: Efficient Lifelong Memory for LLM Agents*. arXiv:2601.02553. https://github.com/aiming-lab/SimpleMem
Maharana, A., et al. (2024). *Evaluating Very Long-Term Conversational Memory of LLM Agents* (LoCoMo). arXiv:2402.17753. https://snap-research.github.io/locomo/
McClelland, J. L., McNaughton, B. L., & O'Reilly, R. C. (1995). *Psychological Review*, 102(3), 419–457.
Kumaran, D., Hassabis, D., & McClelland, J. L. (2016). *Trends in Cognitive Sciences*, 20(7), 512–534.