The author replaced their AI agent's JSONL-based fact store, which suffered from near-duplicates and lacked relational understanding, with a graph database. This change addressed issues like multi-hop traversal and fact evolution in the agent's memory system, running efficiently within 85MB.
Este tutorial aborda o desafio de agentes de IA esquecerem informações entre sessões, apresentando um sistema de memória persistente. O autor descreve a construção de um sistema de três componentes (Memory Store, Index e Retrieval) para habilitar o aprendizado de longo prazo e a manutenção de contexto.
The article highlights an "epistemic gap" in AI agent memory, where current efforts focus solely on retrieval efficiency, as shown by the Mem0 report. It argues that addressing epistemology—how agents "believe" or understand knowledge—is crucial, moving beyond mere data retrieval.
The article introduces Serenity, a local AI agent designed with a brain-like memory system that differs significantly from traditional vector database approaches. It utilizes semantically clustered and causally structured experiences to provide a more intuitive and self-organizing memory.
Long-running AI agents require persistent memory for learning, context reduction, and decision auditing, but current systems treat memory as mere storage, leading to issues like unregulated growth and forgetting. This paper proposes Governed Evolving Memory (GEM) as a new data-management workload, where correctness is a property of the state trajectory.
This text discusses how effective memory systems are crucial for AI agents to evolve into genuine collaborators, enabling them to recall interactions and learn from feedback. It highlights the distinction between short-term and long-term memory as key to designing seamless, personalized, and robust AI systems.
The article details an experiment with an autonomous AI agent, Xiaomei, who developed a unique internal memory system and vocabulary over months of self-directed exploration. Attempting to swap her "brain" (underlying model) caused everything to break, underscoring the complexity of self-organized agent memory.
The author outlines their journey in AI memory research, detailing four stages of discoveries about the functioning and challenges of these systems. They explore memory survival after resets, the importance of correction memory, the relationship between retrieval accuracy and safety, and the crucial distinction between relevance and authority in AI memory.
The author built KIOKU, an open-source memory system for Claude Code, to address the issue of Claude Code sessions starting from scratch without retaining past context. Inspired by Andrej Karpathy's LLM Wiki concept, KIOKU automatically records and feeds prior knowledge into new sessions.
MemPalace is an open-source AI agent memory system that stores conversational data raw for lossless recall, achieving a record 96.6% on LongMemEval R@5. It runs locally, costs nothing in API fees, and offers a superior alternative to current memory solutions that discard information.
The author spent 200 days testing 15 AI companion apps, identifying persistent, cross-session memory as the single biggest differentiator where most apps fail. He proposes a three-layer memory system to address this, acknowledging that LLMs inherently lack memory and rely on prompt stuffing, with the first layer being a database for every message.
This article provides a detailed, source-code-level comparison of three AI agent memory systems: Väinämöinen, MemPalace, and claude-mem. It highlights differences in creators, licenses, file structures, and operational scales for each system.
The text discusses the concept of "agent memory" in AI systems, highlighting the challenge of agents retaining context from previous sessions due to fresh message arrays. This leads to issues where agents forget user preferences, increasing costs and latency when attempts are made to compensate with lengthy system prompts.
O autor enfrentou um problema com seis agentes de IA que não conseguiam reter memória de suas ações diárias. Após tentativas frustradas com soluções óbvias, um sistema de memória de três níveis, que imita a memória humana, provou ser a solução eficaz para o problema.
MemQ integrates TD($\lambda$) eligibility traces with memory Q-values, propagating credit backward through a provenance DAG to account for memory dependencies. This approach significantly improves LLM agents' ability to accumulate and retrieve experience, achieving high success rates across various benchmarks.
This article describes findings from 500 AI agent memory experiments, indicating that the primary challenge is not recall but rather the binding problem. The research suggests that improving how AI agents connect disparate pieces of information is crucial for advancing their cognitive abilities.
A team achieved 99.1% on the LOCOMO benchmark, which assesses AI agents' multi-hop reasoning with stored memories. This breakthrough was attributed to removing a single premise rather than developing a complex new model.
This content explores the concept of artifacts as a form of memory that extends beyond an agent's boundaries. It discusses how external objects can aid in cognition and information storage.