Thuật ngữ AI

Semantic Memory

Structured knowledge base storing general facts and relationships about the environment, independent of specific experiences, to guide the agent's reasoning and planning.

Read-Write Memory

Retrieval Function

Mechanism, often attention-based, that calculates the relevance of each stored item in memory relative to the agent's current state or query to select information to read.

Gated Recurrent Unit with Memory (GRU-M)

Variant of GRUs where internal gates control not only the information flow in the hidden state but also read and write operations in a dedicated external memory.

Reflective Agent

Agent that, before acting, performs a sequence of 'reasoning' or 'reflection' steps by reading and writing to its memory to plan a better strategy, especially in complex environments.

Content-Addressable Memory

Type of memory where data is retrieved based on its content or similarity with a query, rather than by a fixed physical address, enabling associative search.

State Inference

Process by which an agent in a POMDP updates its belief about the hidden state of the environment based on the history of observations, actions, and rewards.

Working Memory

Component of external memory with limited capacity used to temporarily store information relevant to the current task, facilitating reasoning and immediate decision-making.

Relational Memory Core (RMC)

Memory module using a multi-head attention mechanism to model relationships between different memory locations, allowing more complex reasoning on stored information.

Trajectory Memory

Explicit storage of complete or partial trajectories (sequences of states, actions, rewards) to allow the agent to re-evaluate its past experiences or imitate successful behaviors.