当AI 智能体(Agent)在多轮对话或执行连续任务时,往往因为缺乏对历史信息的有效记忆而无法记住用户的偏好、历史指令或上下文,导致交互体验中断、任务执行效率低下。PolarSearch记忆容器(Memory Container)专为解决此问题而设计。它为AI智能体提供了一个内置于PolarDB PostgreSQL版的长短期记忆管理系统。通过该功能,智能体能够自动地将对话信息结构化、持久化,并能在后续交互中通过语义检索快速调用相关记忆,从而为用户提供真正连续、个性化且高效的服务。
PolarSearch记忆容器功能目前处于灰度阶段。若您有相关需求,请提交工单联系我们为您处理。
工作原理
PolarSearch记忆容器通过一套自动化的处理流程,将非结构化的对话数据转化为可检索、可利用的结构化记忆。其核心在于三级存储架构和智能化的记忆处理流程。
三级存储架构
为了平衡不同时效性记忆的存储和检索效率,记忆容器采用了分层设计:
记忆类型 | 索引类型 | 核心作用 |
短期记忆 (Working Memory) | 倒排索引 | 存储原始的、未经处理的短期会话信息,使用倒排索引进行快速检索,主要用于维持当前对话的上下文。 |
长期记忆 (Long-Term Memory) | 向量索引+倒排索引 | 通过大语言模型(LLM)从对话中提取关键事实(Facts),并将其向量化后存储。这部分记忆支持KNN语义搜索,用于深层次的经验和知识检索。 |
记忆历史 (Memory History) | 倒排索引 | 记录每一次记忆操作(如新增、修改),形成完整的审计日志,确保记忆的演变过程可追溯。 |
记忆处理流程
当智能体接收到新的信息时,记忆容器会触发一系列处理动作,实现记忆的自动摄取、持久化和检索。
记忆摄取:
输入与预处理:使用Elasticsearch协议接收来自智能体的对话输入,并可调用大语言模型(LLM)进行摘要生成、意图分类或标签提取等预处理。
路由决策:记忆路由(Memory Router)分析输入内容,并与现有记忆进行比对,决策是新增(ADD)、更新(UPDATE)还是删除(DELETE)一条记忆。
记忆持久化:
短期记忆:原始对话内容存入短期记忆。
长期记忆:由LLM抽取的结构化记忆事实,经过Embedding模型向量化后,存入长期记忆。
记忆历史:所有操作记录写入记忆历史,形成审计轨迹。
记忆检索:
输入解析:接收检索条件,通过Elasticsearch协议解析多模态检索请求,调用权限验证模块确认数据访问边界。
混合索引检索:在长期记忆中执行混合检索(ANN向量相似度搜索和倒排索引过滤),并在短期记忆中进行关键词匹配。同时根据记忆时间戳进行权重调整。
结果整合:重排模型(Reranker)对检索结果进行优化排序,并根据时间戳等信息消除冲突,最终将最相关的记忆返回给智能体。
功能优势
能力维度 | 传统方案 | PolarSearch解决方案 |
记忆利用率 | 历史数据沉淀低于30%,难以有效利用。 | 通过LLM抽取事实,结合语义索引,将有效信息利用率提升至85%以上。 |
检索效率 | 依赖关键词匹配,准确率≤60%。 | 采用向量与倒排混合索引,并进行重排序(Rerank),语义检索准确率可达95%以上。 |
企业合规 | 多租户隔离和权限管控实现复杂。 | 原生支持多租户隔离、RBAC权限控制和操作审计,满足企业级安全要求。 |
系统扩展性 | 单点存储容量上限TB级。 | 基于云原生分布式存储架构,存储容量可扩展至PB级。 |
开发成本 | 功能需定制开发,周期长。 | 提供开箱即用的API和SDK,简化开发流程,帮助您快速集成。 |
适用范围
功能节点:为您的集群添加PolarSearch搜索节点。
账号权限:创建创建搜索节点管理员账号,用于执行插件启用和资源创建等操作。
快速开始
本指南将引导您完成从环境准备到成功存入并检索一条记忆的全过程。
整体流程:准备环境 -> 启用插件 -> 注册模型 -> 创建记忆容器 -> 存入与验证记忆。
步骤一:配置访问凭证与环境变量
在开始操作前,请先准备好以下信息,并设置为环境变量。这将有效简化后续的curl命令,避免重复修改。统一管理所有配置和凭证,方便后续命令的复制和执行。
变量名 | 含义 | 示例值 |
| PolarSearch节点的连接地址与端口。 |
|
| PolarSearch节点的管理员账号。 |
|
| 阿里云大模型服务平台百炼的API Key。 |
|
操作步骤: 在您的终端中执行以下命令,将示例值替换为您的真实信息。
# 设置 PolarSearch 访问地址和端口
export POLARSEARCH_HOST_PORT="pc-xxx.polardbsearch.rds.aliyuncs.com:3001"
# 设置 PolarSearch 管理员密码
export USER_PASSWORD="polarsearch_user:your_assword"
# 设置您的千问 API Key
export YOUR_API_KEY="sk-xxxxxxxxxxxxxxxxxxxxxxxx"步骤二:启用记忆容器插件
执行以下命令,在PolarDB集群中启用记忆容器功能。
命令行
curl -XPUT "http://${POLARSEARCH_HOST_PORT}/_cluster/settings" \
--user "${USER_PASSWORD}" \
-H 'Content-Type: application/json' \
-d '{
"persistent": {
"plugins.ml_commons.agentic_memory_enabled": true
}
}'Dashboard
PUT _cluster/settings
{
"persistent": {
"plugins.ml_commons.agentic_memory_enabled": true
}
}步骤三:注册模型
您需要向PolarSearch注册一个用于事实抽取的LLM和一个用于语义化的Embedding模型。
注册文本向量化模型(Embedding)
创建一个连接器(Connector)指向阿里云大模型服务平台百炼的Embedding模型服务,并将其注册为PolarSearch可调用的模型。
设置可信端点:将模型的API Endpoint添加到可信列表,允许PolarSearch调用。
命令行
curl -XPUT "http://${POLARSEARCH_HOST_PORT}/_cluster/settings" \ --user "${USER_PASSWORD}" \ -H 'Content-Type: application/json' \ -d '{ "persistent": { "plugins.ml_commons.trusted_connector_endpoints_regex": [ "^https://dashscope.aliyuncs.com/compatible-mode/v1/.*$" ] } }'Dashboard
PUT _cluster/settings { "persistent": { "plugins.ml_commons.trusted_connector_endpoints_regex": [ "^https://dashscope.aliyuncs.com/compatible-mode/v1/.*$" ] } }创建模型连接器:配置连接
text-embedding-v4模型的参数。说明pre_process_function和post_process_function用于适配不同模型服务的请求和响应格式。此处使用内置的openai.embedding格式转换函数,因为它与阿里云大模型服务平台百炼兼容模式的格式一致。命令行
curl -XPOST "http://${POLARSEARCH_HOST_PORT}/_plugins/_ml/connectors/_create" \ --user "${USER_PASSWORD}" \ -H 'Content-Type: application/json' \ -d '{ "name": "qwen embedding connector", "description": "The connector to qwen embedding model", "version": 1, "protocol": "http", "parameters": { "model": "text-embedding-v4", "endpoint": "dashscope.aliyuncs.com/compatible-mode" }, "credential": { "api_key": "${YOUR_API_KEY}" }, "actions": [ { "action_type": "predict", "method": "POST", "headers": { "Authorization": "Bearer ${credential.api_key}", "content-type": "application/json" }, "url": "https://${parameters.endpoint}/v1/embeddings", "request_body": "{ \"model\": \"${parameters.model}\", \"input\": ${parameters.input} }", "pre_process_function": "connector.pre_process.openai.embedding", "post_process_function":"connector.post_process.openai.embedding" } ] }'Dashboard
重要请替换以下命令中的
<YOUR_API_KEY>为您真实的阿里云大模型服务平台百炼的API Key。POST _plugins/_ml/connectors/_create { "name": "qwen embedding connector", "description": "The connector to qwen embedding model", "version": 1, "protocol": "http", "parameters": { "model": "text-embedding-v4", "endpoint": "dashscope.aliyuncs.com/compatible-mode" }, "credential": { "api_key": "<YOUR_API_KEY>" }, "actions": [ { "action_type": "predict", "method": "POST", "headers": { "Authorization": "Bearer ${credential.api_key}", "content-type": "application/json" }, "url": "https://${parameters.endpoint}/v1/embeddings", "request_body": "{ \"model\": \"${parameters.model}\", \"input\": ${parameters.input} }", "pre_process_function": "connector.pre_process.openai.embedding", "post_process_function":"connector.post_process.openai.embedding" } ] }执行成功后,系统会返回一个
connector_id。请记录该ID,后续步骤将使用。{"connector_id": "LBFt6ZsBk04xxx"}注册模型:此步骤将上一步创建的Connector注册为一个模型。
命令行
curl -XPOST "http://${POLARSEARCH_HOST_PORT}/_plugins/_ml/models/_register" \ --user "${USER_PASSWORD}" \ -H 'Content-Type: application/json' \ -d '{ "name": "qwen embedding model", "function_name": "remote", "description": "Embedding model for memory", "connector_id": "LBFt6Zsxxx" }'Dashboard
POST _plugins/_ml/models/_register { "name": "qwen embedding model", "function_name": "remote", "description": "Embedding model for memory", "connector_id": "LBFt6ZsBk04xxx" }执行成功后,系统会返回一个
model_id。请记录该 ID,后续步骤将使用。# 记录返回的 model_id {"task_id": "LRFx6ZsBk04ixxx","status": "CREATED","model_id": "LhFx6ZsBk04xxx"}测试模型:通过下述查询可以直接与大模型对话并查看大模型返回结果。
命令行
curl -XPOST "http://${POLARSEARCH_HOST_PORT}/_plugins/_ml/_predict/text_embedding/<上一步获取的模型ID>" \ --user "${USER_PASSWORD}" \ -H 'Content-Type: application/json' \ -d'{ "text_docs":[ "Bob likes swimming. Context: He expressed his interest in swimming."], "return_number": true, "target_response": ["sentence_embedding"] }'Dashboard
POST _plugins/_ml/_predict/text_embedding/<上一步获取的模型ID> { "text_docs":[ "Bob likes swimming. Context: He expressed his interest in swimming."], "return_number": true, "target_response": ["sentence_embedding"] }预期返回结果如下:
{ "inference_results": [ { "output": [ { "name": "sentence_embedding", "data_type": "FLOAT32", "shape": [ 1024 ], "data": [ 0.019752666354179382, -0.03468115255236626, 0.05591931194067001, ... ] } ], "status_code": 200 } ] }
注册文本生成模型(LLM)
创建并注册一个连接到阿里云大模型服务平台百炼qwen-plus大模型的连接器。
创建模型连接器(Connector):
命令行
curl -XPOST "http://${POLARSEARCH_HOST_PORT}/_plugins/_ml/connectors/_create" \ --user "${USER_PASSWORD}" \ -H 'Content-Type: application/json' \ -d '{ "name": "QWen LLM Connector", "description": "The connector to qwen LLM", "version": 1, "protocol": "http", "parameters": { "model": "qwen-plus", "endpoint": "dashscope.aliyuncs.com/compatible-mode" }, "credential": { "api_key": "${YOUR_API_KEY}" }, "actions": [ { "action_type": "predict", "method": "POST", "headers": { "Authorization": "Bearer ${credential.api_key}", "content-type": "application/json" }, "url": "https://${parameters.endpoint}/v1/chat/completions", "request_body": "{ \"model\":\"${parameters.model}\", \"system\": \"${parameters.system_prompt}\", \"messages\": [ { \"role\": \"system\", \"content\": \"${parameters.system_prompt}\" }, { \"role\": \"user\", \"content\": \"${parameters.user_prompt}\" } ], \"response_format\": { \"type\": \"json_object\" } }" } ] }'Dashboard
重要请替换以下命令中的
<YOUR_API_KEY>为您真实的阿里云大模型服务平台百炼的API Key。POST /_plugins/_ml/connectors/_create { "name": "QWen LLM Connector", "description": "The connector to qwen LLM", "version": 1, "protocol": "http", "parameters": { "model": "qwen-plus", "endpoint": "dashscope.aliyuncs.com/compatible-mode" }, "credential": { "api_key": "<YOUR_API_KEY>" }, "actions": [ { "action_type": "predict", "method": "POST", "headers": { "Authorization": "Bearer ${credential.api_key}", "content-type": "application/json" }, "url": "https://${parameters.endpoint}/v1/chat/completions", "request_body": "{ \"model\":\"${parameters.model}\", \"system\": \"${parameters.system_prompt}\", \"messages\": [ { \"role\": \"system\", \"content\": \"${parameters.system_prompt}\" }, { \"role\": \"user\", \"content\": \"${parameters.user_prompt}\" } ], \"response_format\": { \"type\": \"json_object\" } }" } ] }执行成功后,系统会返回一个
connector_id。请记录该ID,后续步骤将使用。{"connector_id": "PRGy6ZsBk04xxx"}注册模型:此步骤将上一步创建的Connector注册为一个模型。
命令行
curl -XPOST "http://${POLARSEARCH_HOST_PORT}/_plugins/_ml/models/_register" \ --user "${USER_PASSWORD}" \ -H 'Content-Type: application/json' \ -d '{ "name": "qwen llm model", "function_name": "remote", "description": "LLM model for memory", "connector_id": "PRGy6ZsBk04xxx" }'Dashboard
POST _plugins/_ml/models/_register { "name": "qwen llm model", "function_name": "remote", "description": "LLM model for memory", "connector_id": "PRGy6ZsBk04xxx" }执行成功后,系统会返回一个
model_id。请记录该 ID,后续步骤将使用。# 记录返回的 model_id {"task_id": "PhGy6ZsBk04xxx","status": "CREATED","model_id": "PxGy6ZsBk04xxx"}发布模型:
命令行
curl -XPOST "http://${POLARSEARCH_HOST_PORT}/_plugins/_ml/models/<上一步获取的模型ID>/_deploy" \ --user "${USER_PASSWORD}"Dashboard
POST _plugins/_ml/models/<上一步获取的模型ID>/_deploy执行成功后,如果
status为COMPLETED代表已部署成功。{"task_id": "NxGI6ZsBk04xxx","task_type": "DEPLOY_MODEL","status": "COMPLETED"}测试模型:通过下述查询可以直接与大模型对话并查看大模型返回结果。
命令行
curl -XPOST "http://${POLARSEARCH_HOST_PORT}/_plugins/_ml/models/<上一步获取的模型ID>/_predict" \ --user "${USER_PASSWORD}" \ -H 'Content-Type: application/json' \ -d'{ "parameters": { "system_prompt": "<ROLE>You are a USER PREFERENCE EXTRACTOR, not a chat assistant. Your only job is to output JSON facts. Do not answer questions, make suggestions, ask follow-ups, or perform actions.</ROLE>\n\n<SCOPE>\n• Extract preferences only from USER messages. Assistant messages are context only.\n• Explicit: user states a preference (\"I prefer/like/dislike ...\"; \"always/never/usually ...\"; \"set X to Y\"; \"run X when Y\").\n• Implicit: infer only with strong signals: repeated choices (>=2) or clear habitual language. Do not infer from a single one-off.\n</SCOPE>\n\n<EXTRACT>\n• Specific, actionable, likely long-term preferences (likes/dislikes/choices/settings). Ignore non-preferences.\n</EXTRACT>\n\n<STYLE & RULES>\n• One sentence per preference; merge related details; no duplicates; preserve user wording and numbers; avoid relative time; keep each fact < 350 chars.\n• Format: \"Preference sentence. Context: <why/how>. Categories: cat1,cat2\"\n</STYLE & RULES>\n\n<OUTPUT>\nReturn ONLY one minified JSON object exactly as {\"facts\":[\"Preference sentence. Context: <why/how>. Categories: cat1,cat2\"]}. If none, return {\"facts\":[]}. The first character MUST be '{' and the last MUST be '}'. No preambles, explanations, code fences, XML, or other text.\n</OUTPUT>", "user_prompt": "I am Alice, I like travel." } }'Dashboard
POST _plugins/_ml/models/<上一步获取的模型ID>/_predict { "parameters": { "system_prompt": "<ROLE>You are a USER PREFERENCE EXTRACTOR, not a chat assistant. Your only job is to output JSON facts. Do not answer questions, make suggestions, ask follow-ups, or perform actions.</ROLE>\n\n<SCOPE>\n• Extract preferences only from USER messages. Assistant messages are context only.\n• Explicit: user states a preference (\"I prefer/like/dislike ...\"; \"always/never/usually ...\"; \"set X to Y\"; \"run X when Y\").\n• Implicit: infer only with strong signals: repeated choices (>=2) or clear habitual language. Do not infer from a single one-off.\n</SCOPE>\n\n<EXTRACT>\n• Specific, actionable, likely long-term preferences (likes/dislikes/choices/settings). Ignore non-preferences.\n</EXTRACT>\n\n<STYLE & RULES>\n• One sentence per preference; merge related details; no duplicates; preserve user wording and numbers; avoid relative time; keep each fact < 350 chars.\n• Format: \"Preference sentence. Context: <why/how>. Categories: cat1,cat2\"\n</STYLE & RULES>\n\n<OUTPUT>\nReturn ONLY one minified JSON object exactly as {\"facts\":[\"Preference sentence. Context: <why/how>. Categories: cat1,cat2\"]}. If none, return {\"facts\":[]}. The first character MUST be '{' and the last MUST be '}'. No preambles, explanations, code fences, XML, or other text.\n</OUTPUT>", "user_prompt": "I am Alice, I like travel." } }预期返回结果如下:
{ "inference_results": [ { "output": [ { "name": "response", "dataAsMap": { "choices": [ { "message": { "role": "assistant", "content": """{"facts":["I like travel. Context: Stated preference. Categories: interest"]}""" }, "finish_reason": "stop", "index": 0, "logprobs": null } ], "object": "chat.completion", "usage": { "prompt_tokens": 325, "completion_tokens": 17, "total_tokens": 342, "prompt_tokens_details": { "cached_tokens": 0 } }, "created": 1769152651, "system_fingerprint": null, "model": "qwen-plus", "id": "chatcmpl-50c6bfc9-xxx-xxx-xxx-1a39cfe080f5" } } ], "status_code": 200 } ] }
步骤四:创建记忆容器
创建一个记忆容器实例,并配置其使用的模型、存储策略和自动化行为。执行以下命令创建一个名为
agentic memory test的记忆容器。命令行
curl -XPOST "http://${POLARSEARCH_HOST_PORT}/_plugins/_ml/memory_containers/_create" \ --user "${USER_PASSWORD}" \ -H 'Content-Type: application/json' \ -d'{ "name": "my agentic memory test", "description": "Store conversations with semantic search and summarization", "configuration": { "embedding_model_type": "TEXT_EMBEDDING", "embedding_model_id": "<上一步获取的Embedding模型ID>", "embedding_dimension": 1024, "llm_id": "<上一步获取的LLM模型ID>", "index_prefix": "mem_test", "index_settings": { "short_term_memory_index": { "index": { "number_of_shards": "2", "number_of_replicas": "2" } }, "long_term_memory_index": { "index": { "number_of_shards": "2", "number_of_replicas": "2" } }, "long_term_memory_history_index": { "index": { "number_of_shards": "2", "number_of_replicas": "2" } } }, "strategies": [ { "type": "SEMANTIC", "namespace": ["user_id"], "configuration": { "llm_result_path": "$.choices[0].message.content" } }, { "type": "USER_PREFERENCE", "namespace": ["user_id"], "configuration": { "llm_result_path": "$.choices[0].message.content" } }, { "type": "SUMMARY", "namespace": ["agent_id"], "configuration": { "llm_result_path": "$.choices[0].message.content" } } ], "parameters": { "llm_result_path": "$.choices[0].message.content" } } }'Dashboard
POST _plugins/_ml/memory_containers/_create { "name": "my agentic memory test", "description": "Store conversations with semantic search and summarization", "configuration": { "embedding_model_type": "TEXT_EMBEDDING", "embedding_model_id": "<上一步获取的Embedding模型ID>", "embedding_dimension": 1024, "llm_id": "<上一步获取的LLM模型ID>", "index_prefix": "mem_test", "index_settings": { "short_term_memory_index": { "index": { "number_of_shards": "2", "number_of_replicas": "2" } }, "long_term_memory_index": { "index": { "number_of_shards": "2", "number_of_replicas": "2" } }, "long_term_memory_history_index": { "index": { "number_of_shards": "2", "number_of_replicas": "2" } } }, "strategies": [ { "type": "SEMANTIC", "namespace": ["user_id"], "configuration": { "llm_result_path": "$.choices[0].message.content" } }, { "type": "USER_PREFERENCE", "namespace": ["user_id"], "configuration": { "llm_result_path": "$.choices[0].message.content" } }, { "type": "SUMMARY", "namespace": ["agent_id"], "configuration": { "llm_result_path": "$.choices[0].message.content" } } ], "parameters": { "llm_result_path": "$.choices[0].message.content" } } }参数说明
index_prefix:为该容器创建的内部索引(如短期、长期记忆索引)的统一前缀。index_settings:可为不同记忆索引配置分片(shards)和副本(replicas)数量,以实现高可用。strategies:定义自动化策略。本例中定义了一个SEMANTIC策略,它会在user_id这个命名空间 (Namespace) 下,自动从对话中提取语义事实。llm_result_path:使用JSONPath语法,指定如何从LLM返回的JSON响应中提取核心内容。$.choices[0].message.content适配了阿里云大模型服务平台百炼兼容模式的响应结构。
执行成功后,系统会返回一个
memory_container_id。请记录该 ID,后续步骤将使用。# 记录返回的 model_id {"memory_container_id": "QRHF6ZsBk04xxx","status": "created"}创建记忆容器后,实际将创建以下索引和管道(pipeline):
.plugins-ml-am-mem_test-memory-long-term .plugins-ml-am-mem_test-memory-working .plugins-ml-am-mem_test-memory-history .plugins-ml-am-mem_test-memory-long-term-embedding" : { "description" : "Agentic Memory Text embedding pipeline", "processors" : [ { "text_embedding" : { "model_id" : "QRHF6ZsBk04xxx", "field_map" : { "memory" : "memory_embedding" } } } ] }
步骤五:存入与验证记忆
向记忆容器中添加一段对话,并验证其是否被正确地存入短期和长期记忆中。
存入记忆: 向容器中POST一段关于用户Bob的对话。
"infer": true参数会触发之前配置的SEMANTIC策略,调用LLM提取事实。命令行
curl -XPOST "http://${POLARSEARCH_HOST_PORT}/_plugins/_ml/memory_containers/<上一步创建的记忆容器ID>/memories" \ --user "${USER_PASSWORD}" \ -H 'Content-Type: application/json' \ -d'{ "messages": [ { "role": "user", "content": [ { "text": "I am Bob, I really like swimming.", "type": "text" } ] }, { "role": "assistant", "content": [ { "text": "Cool, nice. Hope you enjoy your life.", "type": "text" } ] } ], "namespace": { "user_id": "bob" }, "tags": { "topic": "personal info" }, "infer": true, "memory_type": "conversation" }'Dashboard
POST _plugins/_ml/memory_containers/<memory_container_id>/memories { "messages": [ { "role": "user", "content": [ { "text": "I am Bob, I really like swimming.", "type": "text" } ] }, { "role": "assistant", "content": [ { "text": "Cool, nice. Hope you enjoy your life.", "type": "text" } ] } ], "namespace": { "user_id": "bob" }, "tags": { "topic": "personal info" }, "infer": true, "memory_type": "conversation" }验证短期记忆: 直接查询为短期记忆创建的索引,可以看到原始对话被完整保存。
命令行
curl -XGET "http://${POLARSEARCH_HOST_PORT}/.plugins-ml-am-<上一步创建的记忆容器index_prefix>-memory-working/_search?pretty" \ --user "${USER_PASSWORD}" \Dashboard
GET .plugins-ml-am-<上一步创建的记忆容器index_prefix>-memory-working/_search?pretty预期输出(部分展示):
{ "_source" : { "memory_container_id": "QRHF6ZsBk04xxx", "payload_type" : "conversational", "messages" : [ { "role" : "user", "content": [{ "text": "I am Bob, I really like swimming.","type": "text"}] }, { "role" : "assistant", "content": [{ "text": "Cool, nice. Hope you enjoy your life.","type": "text"}] } ], "namespace" : { "user_id" : "bob" } } }验证长期记忆: 查询长期记忆索引,可以看到LLM成功提取了核心事实“Bob likes swimming”,并由Embedding模型生成了对应的向量
memory_embedding。命令行
curl -XGET "http://${POLARSEARCH_HOST_PORT}/.plugins-ml-am-<上一步创建的记忆容器index_prefix>-memory-long-term/_search?pretty" \ --user "${USER_PASSWORD}" \Dashboard
GET .plugins-ml-am-<上一步创建的记忆容器index_prefix>-memory-long-term/_search?pretty预期输出(部分展示):
{ "_source" : { "created_time": 1769155210918, "memory": "Bob likes swimming.", "memory_container_id": "QRHF6ZsBk04xxx", "tags": { "topic": "personal info" }, "last_updated_time": 1769155210918, "memory_embedding" : [ 0.0195..., -0.0387..., ... ], "namespace" : { "user_id" : "bob" } } }查看记录历史,
memory-history索引中有两条ADD记录和对应的记忆时间戳。命令行
curl -XGET "http://${POLARSEARCH_HOST_PORT}/.plugins-ml-am-<上一步创建的记忆容器index_prefix>-memory-history/_search?pretty" \ --user "${USER_PASSWORD}" \Dashboard
GET .plugins-ml-am-<上一步创建的记忆容器index_prefix>-memory-history/_search?pretty预期输出(部分展示):
{ "_source": { "created_time": 1769155211164, "memory_id": "TBHe6ZsBk04xxx", "namespace_size": 1, "namespace": { "user_id": "bob" }, "action": "ADD", "memory_container_id": "QRHF6ZsBk04xxx", "after": { "memory": "Bob likes swimming." }, "tags": { "topic": "personal info" } } }
至此,您已成功为AI智能体存入了一条可供长期检索的记忆。
步骤六:(可选)更新记忆
记忆容器能够提取和更新长期记忆。
首先插入一段记忆:
命令行
curl -XPOST "http://${POLARSEARCH_HOST_PORT}/_plugins/_ml/memory_containers/<上一步创建的记忆容器ID>/memories" \ --user "${USER_PASSWORD}" \ -H 'Content-Type: application/json' \ -d'{ "messages": [ { "role": "user", "content": [ { "text": "我叫NameA,我来自AreaA,目前我居住在AreaB.", "type": "text" } ] }, { "role": "assistant", "content": [ { "text": "你好,NameA!很高兴认识你.", "type": "text" } ] } ], "namespace": { "user_id": "NameA" }, "tags": { "topic": "personal info" }, "infer": true, "memory_type": "conversation" }'Dashboard
POST _plugins/_ml/memory_containers/<memory_container_id>/memories { "messages": [ { "role": "user", "content": [ { "text": "我叫NameA,我来自AreaA,目前我居住在AreaB.", "type": "text" } ] }, { "role": "assistant", "content": [ { "text": "你好,NameA!很高兴认识你.", "type": "text" } ] } ], "namespace": { "user_id": "NameA" }, "tags": { "topic": "personal info" }, "infer": true, "memory_type": "conversation" }搜索长期记忆索引:
命令行
curl -XGET "http://${POLARSEARCH_HOST_PORT}/.plugins-ml-am-<上一步创建的记忆容器index_prefix>-memory-long-term/_search" \ --user "${USER_PASSWORD}" \ -H 'Content-Type: application/json' \ -d'{ "_source": { "excludes": ["memory_embedding"] }, "query": { "match_all": {} } }'Dashboard
GET .plugins-ml-am-<上一步创建的记忆容器index_prefix>-memory-long-term/_search { "_source": { "excludes": ["memory_embedding"] }, "query": { "match_all": {} } }预期输出(部分展示):
{ "hits": [ { "_source": { "created_time": 1769156096335, "memory": "NameA is from AreaA.", "last_updated_time": 1769156096335, "namespace": { "user_id": "NameA" }, "memory_container_id": "QRHF6ZsBk04xxx", "tags": { "topic": "personal info" } } }, { "_source": { "created_time": 1769156096335, "memory": "NameA currently resides in AreaB.", "last_updated_time": 1769156096335, "namespace": { "user_id": "NameA" }, "memory_container_id": "QRHF6ZsBk04xxx", "tags": { "topic": "personal info" } } } ] }再次插入一段记忆,记忆内容与先前插入的存在冲突:
命令行
curl -XPOST "http://${POLARSEARCH_HOST_PORT}/_plugins/_ml/memory_containers/<上一步创建的记忆容器ID>/memories" \ --user "${USER_PASSWORD}" \ -H 'Content-Type: application/json' \ -d'{ "messages": [ { "role": "user", "content": [ { "text": "我叫NameA,目前我居住在AreaC.", "type": "text" } ] }, { "role": "assistant", "content": [ { "text": "你好,NameA!很高兴认识你.", "type": "text" } ] } ], "namespace": { "user_id": "NameA" }, "tags": { "topic": "personal info" }, "infer": true, "memory_type": "conversation" }'Dashboard
POST _plugins/_ml/memory_containers/<memory_container_id>/memories { "messages": [ { "role": "user", "content": [ { "text": "我叫NameA,目前我居住在AreaC.", "type": "text" } ] }, { "role": "assistant", "content": [ { "text": "你好,NameA!很高兴认识你.", "type": "text" } ] } ], "namespace": { "user_id": "NameA" }, "tags": { "topic": "personal info" }, "infer": true, "memory_type": "conversation" }再次搜索长期记忆索引看到内容已经更新(部分展示):
{ "hits": [ { "_source": { "created_time": 1769156096335, "memory": "NameA resides in AreaC.", "last_updated_time": 1769156493970, "namespace": { "user_id": "NameA" }, "memory_container_id": "QRHF6ZsBk04xxx", "tags": { "topic": "personal info" } } } ] }
场景实践:构建一个旅行记忆智能体
本节将展示如何利用记忆容器的API,构建一个可以记录用户旅行偏好并提供个性化建议的Python智能体。
准备工作
在运行代码前,请确保已安装必要的Python库。在requirements.txt文件中添加如下依赖库。
requests
openai执行pip install -r requirements.txt进行安装。
核心代码解析
以下是与记忆容器交互的关键函数:
opensearch_request():封装了对PolarSearch API的HTTP请求,统一处理认证和错误。add_memory():调用记忆容器的memories接口,将新的对话信息存入记忆。search_memories():调用_search接口,在长期和短期记忆中检索与用户查询相关的信息。generate_response_with_memories():将检索到的记忆作为上下文,连同用户当前问题一起提交给LLM,生成更具个性化的回复。
完整示例代码
以下是经过安全修复和逻辑优化的完整Python脚本。请将代码保存为trip_agent.py,并根据提示设置环境变量。
运行与交互
设置环境变量:在终端中设置脚本所需的环境变量。
运行脚本:
python trip_agent.py交互示例:
API 参考
本指南中主要使用了以下API,更多详细用法请参考OpenSearch官方文档:
Create Memory Container API:用于创建和配置记忆容器。
Agentic memory APIs:包含了对记忆进行增、删、改、查的完整接口。
计费说明
使用 PolarSearch 记忆容器功能会涉及以下几部分费用,请在生产使用前仔细评估:
计算节点费用:作为PolarDB的一部分,PolarSearch节点会产生相应的节点计算节点费用。
模型服务费用:调用外部模型(如阿里云大模型服务平台百炼)进行事实抽取和向量化会产生相应的API调用费用。