long-term-memory

The paper introduces REALM, a framework for long-term memory in LLM agents that uses retrieval-driven reconsolidation to autonomously organize memories into a cognitive graph, achieving improved performance on long-term memory benchmarks.

CueMem is a cue-guided framework for long-term conversational memory that reconstructs dialogue context using retrieval cues to improve performance over existing baselines in LLM-powered agents.

The article proposes a Markdown-based memory system for AI agents that uses simple file storage and editing rules to outperform complex memory runtimes, ensuring accurate and sourced fact retrieval.

ROAM introduces a relation-guided framework for managing atomic memories in AI agents, improving answer accuracy by up to 29.8 percentage points through semantic classification and memory fusion.

An individual built a local long-term memory system for AI agents using markdown files and a local index, enabling persistent memory across sessions and including tests for false memories, with plans to potentially productize it.

MemoryLACE is a lightweight memory framework for long-term LLM agents that explicitly models the lifecycle of textual evidence to improve reasoning without comprehensive knowledge graphs, achieving higher performance and reduced runtime on benchmarks.

The article discusses challenges and asks for community experiences regarding the breakdown of AI agent memory systems after months in production use.

Garry Tan has released new evaluations for gbrain, an open-source retrieval layer for AI agents, showcasing state-of-the-art performance in memory reading and writing without LLM-in-loop retrieval.

Ring Forcing is an autoregressive video diffusion framework that enhances long-term memory for precise minute-long video coherence through ring-structured training, history compression, and sparse rotary embeddings.

The article discusses the complexity of Pi's long-term memory, noting that Pi natively focuses on session and context management, while the community offers multiple memory implementations akin to Hermes. Four projects are recommended for implementing AI agent memory functionality.

The paper introduces MemUse, a benchmark that shows direct QA accuracy for conversational memory does not predict user satisfaction, whereas natural integration of prior context does, revealing a large gap between recall and conversational use.

CABLE is a plug-in augmentation that extends LLM memory retrieval by constructing complementary antecedent-based links to surface implicit supporting evidence, evaluated on benchmarks with higher mean scores across multiple systems.

ArborMem introduces an online memory framework for large language models that represents conversations as a navigable forest of interaction states, outperforming baselines on memory benchmarks and introducing BranchMemEval as a new diagnostic benchmark.

The study compares MemoryLake, a structured memory backend, with other systems on the MemoryArena benchmark, finding it has higher success rates in some tasks but not establishing a clear advantage across all domains.

The project introduces long-term personality and memory persistency for AI companion agents, featuring mechanisms like evolving personalities, grudge-holding, and associative memory web.

Presents EvoGraph-Mem, a failure-aware editable graph memory framework for long-term language agents that tracks positive/negative evidence and activation states for insights, enabling memory maintenance through utility-aware retrieval and graph-level editing.

LycheeMemory V2 introduces semantic segment-level consolidation for LLM agent memory, reducing construction costs while improving or maintaining accuracy on LoCoMo and LongMemEval benchmarks.

Introduces Controlled Memory Interference (CMI), a diagnostic framework for studying how LLM agent memory evolves under different memory relationships, revealing that relationship-specific interference suppresses update plasticity and that interference-aware training improves valid update distinction.

MobileMem introduces a benchmark and framework for on-device AI systems to learn from year-long mobile experiences, focusing on temporal reasoning, knowledge updating, and preference inference.

After leaving Shanda Innovation Institute, Lin Kaisen founded a new company focused on long-term memory agents, and mentioned his experience working under the guidance of Boss Chen.