Spring AI 2.1 stores messages as ordered parts so model reasoning survives replay

Victoria Shinder29 October 2026, 09:202 min read

Spring AI 2.1 rebuilds how a message is represented: content is no longer a string with side lists but a sequence of typed pieces kept in the order the model produced them, the Spring team explains in a post dated 28 October. Each piece — prose, model thinking, a tool invocation — is mapped from what providers call content blocks (Anthropic, Bedrock), parts (Gemini), items (OpenAI Responses) or chunks (Mistral).

The reason is reasoning models. Anthropic's models attach a signature to their thinking, Google's attach signatures to function calls, and OpenAI's Responses API hands back reasoning in encrypted form; each provider expects that data back unchanged and in order on the next request. Dropping or reordering them can mean an HTTP 400 or a model that silently loses its earlier reasoning. Before 2.1, text, tool calls and media sat in separate fields, opaque data lived in provider-specific metadata, streaming often kept only the last fragment of thinking, and the JDBC chat memory stored only text and message type.

MessagePart is a sealed interface with six record implementations. Opaque payloads such as an Anthropic signature are carried as OpaquePayload(provider, kind, data) and replayed only to the provider that produced them; anything a model does not understand becomes an UnknownPart and is passed back as-is. With 2.1.0-RC1, OpenAI Responses, Anthropic, Google GenAI, OpenAI Chat Completions, Ollama, Bedrock Converse, DeepSeek and Mistral produce and consume parts, and a new Session memory persists them. Existing constructors and builders still work, and getText(), getToolCalls() and getMedia() become views over the parts; code that read reasoning through provider-specific types is what most likely needs changes.

What it means

The three rules in the post — keep the order, replay verbatim, persist parts rather than text — apply to any framework talking to reasoning models, not just Spring. Teams storing conversation history as plain text are exactly the ones who will see multi-turn tool use degrade without an obvious error. Note that 2.1 is at release-candidate stage.