Add Context Memory to Your AI in 5 Lines with Spring AI ChatMemory

When an AI chatbot receives a request, it only sees the current turn, so asking "continue" leads to a confused response because the model has no memory of previous interactions. This stateless behavior is a classic multi‑turn dialogue problem.

Spring AI solves the issue with ChatMemory , which stores each exchange and automatically includes the conversation history in subsequent prompts, effectively creating an "AI memory palace".

The ChatMemory stack consists of three layers:

ChatMemoryRepository (storage layer interface)
    ↑
ChatMemory (manager)
    ↑
MessageChatMemoryAdvisor (interceptor registered to ChatClient)

Using Spring Boot style, the entire solution can be implemented in five lines of code inside a controller:

@RestController
@RequestMapping("chat/memory")
public class ChatMemoryController {
    @Resource
    private MiniMaxChatModel miniMaxChatModel;
    @Resource
    private ChatMemory chatMemory; // auto‑configured InMemory instance

    @GetMapping(value = "v1", produces = MediaType.TEXT_EVENT_STREAM_VALUE)
    public Flux<String> streamEvent(@RequestParam("question") String question) {
        // 5 core lines
        ChatClient chatClient = ChatClient.builder(miniMaxChatModel)
            .defaultAdvisors(MessageChatMemoryAdvisor.builder(chatMemory).build())
            .build();
        return chatClient.prompt()
            .user(question)
            .stream()
            .content();
    }
}

The advisor automatically reads historical messages, appends them to the prompt, and writes the new user‑AI exchange back to the memory.

Different storage back‑ends can be chosen:

InMemory – zero‑configuration, suitable for development; loses data on restart.

JDBC – production‑ready; add the spring-ai-starter-model-chat-memory-repository-jdbc dependency and set

spring.ai.chat.memory.repository.jdbc.initialize-schema=embedded

to let Spring create the table.

Redis – high performance with advanced queries; configure spring.ai.chat.memory.redis.host, port, and time-to-live. It offers content search and is about ten times faster than JDBC.

The advisor works as follows:

User request
  ↓
MessageChatMemoryAdvisor (interceptor)
  ├─ Read: chatMemory.get(sessionId) → history
  ├─ Append: history + current question → full prompt
  ↓
Send to AI model
  ↓
AI response
  ↓
MessageChatMemoryAdvisor
  ├─ Write: chatMemory.add(sessionId, userMsg, aiReply)
  ↓
Return to user

Session isolation is achieved by supplying a unique sessionId per user, e.g., String sessionId = user.getId(); and constructing the advisor with that ID.

Because AI models have a token limit (MiniMax typically 32K), memory size must be controlled. Spring AI provides two window implementations: MessageWindowChatMemory.builder().maxMessages(10).build() – keeps the most recent N messages (10‑20 works for simple dialogs). TokenWindowChatMemory.builder().maxTokens(2000).build() – caps the total token count for longer conversations.

Practical tips gathered from experience:

Use maxMessages=10 for simple Q&A, maxMessages=20 for complex dialogs, and consider maxTokens=2000 for very long sessions.

Never share a single session ID across users; always isolate memory per user ID.

Avoid InMemory in production; switch to Redis or JDBC to persist memory across restarts.

Monitor token usage to prevent truncation when the context exceeds the model's limit.

In summary, ChatMemory is a fundamental building block for AI applications, and Spring AI makes its integration as straightforward as typical database operations, with the Advisor mechanism handling memory automatically and session IDs ensuring isolation.