返回新聞
產品展示AI Understanding 簡報

Amazon Bedrock 新增具備記憶感知的代理人,搭配 Agentic Retrieve 以管理知識庫

AWS 推出 Agentic Retrieve,讓 Bedrock 代理人能儲存對話歷史、提取事實,並在會話間重複使用,同時遵守每位使用者的存取控制。

4 min readRead the primary source
Source-provided image accompanying Amazon Bedrock adds memory‑aware agents with Agentic Retrieve for managed knowledge bases
主要來源文件來源記錄
出版商
aws.amazon.com
來源類型
主要文件-我們直接閱讀的官方公告、文件、文件或第一方頁面。
背景60 秒內了解這一點

關鍵術語

記憶體(代理記憶體)
AI 代理程式跨步驟或會話使用儲存的上下文來提高連續性。
API(應用程式介面)
一種軟體系統向另一個系統發送請求並接收回應的結構化方式。
知識庫
用於檢索、支援自動化或基礎回應的精選文件或記錄集合。

發生了什麼事

Amazon Bedrock now supports a memory‑aware agent capability called Agentic Retrieve. By adding a memoryConfiguration to an AgenticRetrieveStream request, developers can have Bedrock automatically store each turn of a conversation, extract durable facts, and pull relevant facts into subsequent queries. The feature works with Bedrock Managed , enabling per‑user ACL filtering and private, isolated memory namespaces. The blog post provides a step‑by‑step walkthrough building a multi‑tenant employee assistant that authenticates users via Amazon Cognito, enforces document access via ACLs, and isolates each user’s long‑term memory within a shared AgentCore memory resource.

Amazon Bedrock’s Managed now includes an optional memoryConfiguration that can be supplied to the AgenticRetrieveStream API. When enabled, the service restores a user’s prior conversation from AgentCore memory, lets the planner decide if a stored fact would improve the answer, retrieves that fact, and incorporates it into the sub‑queries sent to the knowledge base. After generating a response, the turn is saved back to memory.

The blog demonstrates a multi‑tenant employee assistant for a fictional company, AnyCompany. Employees sign in with Amazon Cognito; the Cognito ID token provides a stable sub claim (used as the actorId for memory) and a verified email (used for ACL filtering). Documents are stored in Amazon S3 with a global ACL file that maps S3 prefixes to allowed email addresses.

A single AgentCore memory resource is shared across all users, but isolation is achieved through namespace templates that include the actorId (e.g., "/anycompany/<sub>"). Two extraction strategies are described: a summary strategy that captures any fact mentioned in conversation, and a preference strategy that records explicit user preferences. The planner automatically decides when to invoke memory retrieval based on the current question.

The post also covers required IAM policies, role assumptions with session tags, and how to enforce per‑user memory isolation even if application code is compromised. It explains how to list sessions, retrieve conversation history, and delete a user’s memory records for compliance. Observability is supported via CloudWatch metrics and Bedrock AgentCore tracing.

Finally, the sample code and infrastructure scripts are provided in a GitHub repository, allowing developers to replicate the setup, test the end‑to‑end flow, and tear down resources when finished.

來源詳情: aws.amazon.com

為什麼這很重要

The addition of built‑in memory to Bedrock agents addresses a long‑standing limitation of retrieval‑augmented generation: the inability to retain context beyond a single request. By handling conversation restoration, fact extraction, and memory persistence as a managed service, AWS removes the need for developers to build custom storage layers, reducing engineering effort and potential security gaps. Per‑user isolation and ACL‑aware retrieval also make the feature suitable for enterprise use cases where data privacy and compliance are critical, such as internal HR or sales assistants. This could accelerate adoption of AI‑driven knowledge assistants across large organizations and set a new baseline for “memory‑aware” agents in the cloud AI market.

Memory‑aware agents close the gap between single‑turn retrieval‑augmented generation and truly conversational assistants that can remember user‑specific facts over weeks or months, enabling more natural and personalized interactions.

By offering this as a managed service, AWS eliminates the need for custom databases, vector stores, or bespoke pipelines to persist and retrieve conversation context, lowering development costs and reducing security risks associated with mishandled user data.

Per‑user ACL enforcement ensures that sensitive corporate documents are only returned to authorized employees, addressing compliance concerns that have limited the deployment of AI assistants in regulated environments.

The feature’s integration with existing Bedrock guardrails and observability tools provides enterprises with the auditability required for internal governance and external regulatory scrutiny.

If widely adopted, this could set a new standard for cloud AI platforms, prompting competitors to introduce comparable managed memory capabilities, thereby accelerating the overall maturity of enterprise AI assistants.

Interactive Mechanism

互動機制:它實際上是如何運作的

以互動方式探索這項發展背後的基礎技術。

Agent Lifecycle Stage:
1
User Intent & Planning: "Audit customer refund request #4092 and settle payment."
2
Tool Calling: Emits structured JSON call crm_get_transaction(id='4092').
3
Guardrail & Verification:Paused: High-value action requires human operator sign-off.
4
Final Settlement: Refund recorded, email receipt dispatched, and audit log stored.
Core takeaway: An AI agent is not just a language model—it is a closed loop of planning, tool invocation, and environment feedback. Production systems require self-healing retries and strict human approval guardrails.
互動式概念檢查+10 Points
What is AI? Quiz

A route planner searches possible journeys using explicit rules. What does this illustrate about AI?

接下來看什麼

Watch for developer adoption metrics, pricing details, and any limits on memory size or retention that AWS may publish. Integration with Bedrock guardrails and observability tools will be key for enterprises concerned about compliance and auditability. Competitors may respond with similar managed memory offerings, so the feature could spark a broader shift toward built‑in, privacy‑preserving agent memory in cloud AI services.

Pricing and quota information for AgentCore memory (storage size, retention periods, request rates) to understand cost implications for large‑scale deployments.

Limits on the number of concurrent memory namespaces or the size of extracted records, which could affect scalability for organizations with thousands of users.

Future enhancements such as real‑time memory updates, richer extraction strategies, or tighter integration with third‑party identity providers beyond Cognito.

Customer case studies or benchmark results that demonstrate performance, latency, and accuracy improvements when using memory‑aware retrieval versus stateless approaches.

Responses from other cloud AI providers (e.g., Google Vertex AI, Microsoft Azure AI) that may introduce similar managed memory services, influencing competitive dynamics.

相關指引和測驗

覺得有用嗎?