Long-Term Memory
A component of the Agent memory mechanism referring to information persisted across sessions, such as user preferences and historical interaction patterns. It is typically stored in external databases or vector stores and dynamically retrieved into context to support personalized service.
Core Facts
An agent's long-term memory can be subdivided into three categories: episodic memory, semantic memory, and procedural memory.
Long-term memory is typically implemented with the help of vector databases (such as Chroma, Pinecone, Weaviate, Milvus, etc.) to enable semantic retrieval.
Enterprise-grade Agent systems typically introduce external memory modules (such as vector databases) to store long-term memory, in order to overcome context window limitations.
Timeline (last 90 days)
Every conversation is brand new to AI; it has no true long-term memory and cannot remember the user's preferences and background, and this problem can only be solved with an external database.
The context window is not the same as long-term memory; it is only valid within a single session and is cleared once the conversation ends.
Sandboxing can limit the permission boundary of a single run, but it is difficult to break the chain of errors propagated across tasks through long-term memory.
When told the answer directly, the brain has almost no retrieval or reconstruction work to do, making it difficult for the information to enter long-term memory.
The core contradiction in long-term memory design is: the more that is remembered, the higher the cost and noise, and the less timely the cleanup, the worse the accuracy — the industry currently has no standard answer.
The experience summaries generated during the reflection phase of long-term memory need to be specific enough, ideally including the cause of failure, the corrective approach, and the applicable conditions; otherwise it is difficult to surface useful historical lessons during retrieval.
A knowledge base is static content retrieval that answers what is written in a document, whereas long-term memory focuses on what has been done, what has changed, and what knowledge has been accumulated.
The paper's framework is built on a general agent architecture, describing agent behavior as a combination of cognitive functions organized around long-term memory containing evolved knowledge.
Conversational recall benchmarks such as LoCoMo and LongMemEval measure question-answering ability over conversation history, rather than whether memory has changed the actual behavior of tool-using agents.
An agent's long-term memory can be subdivided into three categories: episodic memory, semantic memory, and procedural memory.
4 more timeline events
All Facts (17)
An agent's long-term memory can be subdivided into three categories: episodic memory, semantic memory, and procedural memory.
80%VerifiedLong-term memory is typically implemented with the help of vector databases (such as Chroma, Pinecone, Weaviate, Milvus, etc.) to enable semantic retrieval.
80%VerifiedEnterprise-grade Agent systems typically introduce external memory modules (such as vector databases) to store long-term memory, in order to overcome context window limitations.
75%VerifiedLangGraph supports two storage modes: short-term memory (single-session context) and long-term memory (cross-session persistent storage).
75%VerifiedMainstream industry solutions for long-term memory include embedding-based vector retrieval, knowledge-graph-based structured storage, and hybrid architectures combining both.
65%VerifiedThe goal of memory poisoning attacks is to embed malicious instructions in an Agent's persistent memory so that it executes unintended operations in future interactions, which differs from the immediate hijacking of prompt injection.
65%VerifiedPersistent memory is typically divided into three levels: short-term working memory, episodic memory, and semantic memory
65%UnverifiedEvery conversation is brand new to AI; it has no true long-term memory and cannot remember the user's preferences and background, and this problem can only be solved with an external database.
50%UnverifiedThe context window is not the same as long-term memory; it is only valid within a single session and is cleared once the conversation ends.
50%UnverifiedSandboxing can limit the permission boundary of a single run, but it is difficult to break the chain of errors propagated across tasks through long-term memory.
50%UnverifiedWhen told the answer directly, the brain has almost no retrieval or reconstruction work to do, making it difficult for the information to enter long-term memory.
50%UnverifiedThe core contradiction in long-term memory design is: the more that is remembered, the higher the cost and noise, and the less timely the cleanup, the worse the accuracy — the industry currently has no standard answer.
50%UnverifiedThe experience summaries generated during the reflection phase of long-term memory need to be specific enough, ideally including the cause of failure, the corrective approach, and the applicable conditions; otherwise it is difficult to surface useful historical lessons during retrieval.
50%UnverifiedA knowledge base is static content retrieval that answers what is written in a document, whereas long-term memory focuses on what has been done, what has changed, and what knowledge has been accumulated.
50%UnverifiedThe paper's framework is built on a general agent architecture, describing agent behavior as a combination of cognitive functions organized around long-term memory containing evolved knowledge.
50%UnverifiedConversational recall benchmarks such as LoCoMo and LongMemEval measure question-answering ability over conversation history, rather than whether memory has changed the actual behavior of tool-using agents.
50%UnverifiedAI has no true long-term memory; every conversation starts fresh, and this can only be addressed by connecting an external database.
50%