RAG Pipeline with in-memory conversation history

[adhikari23]

In Haystack 1.0, we have conversation agent that helps us to use the chat history in the pipeline through a agentic way. https://haystack.deepset.ai/tutorials/24_building_chat_app
However, in Haystack 2.x, I don't see the support of conversation memory agent.
(https://docs.haystack.deepset.ai/docs/migration#agents)

What is the best way to build a pipeline with conversation chat history in Haystack 2.x

[davidsbatista]

hi @adhikari23 it's still not supported we are working on it, but it will soon come in haystack-experimental

[yahya-maged]

So is there is any other way to implement chat history in haystack 2.0? Even if using the ConversationMemory from 1.0 in my haystack 2.0 app

[julian-risch]

Hello @yahya-maged , right now you can try out the InMemoryChatMessageStore with haystack-experimental. ChatMessageRetriever and ChatMessageWriter are open PRs.

[musaabhasan]

For Haystack 2.x I would build conversation memory as an explicit part of the pipeline rather than waiting for it to behave like the 1.x agent abstraction.

A practical layout is:

• keep the authoritative document store separate from chat history
• pass recent turns as a bounded window into a query-rewrite component
• optionally maintain a rolling summary for older turns
• retrieve from the document store using the rewritten standalone question
• send the generator: system instruction, retrieved documents, recent turns, and the summary as separate inputs
• write only the new turn and a small structured summary back to memory

This separation matters because chat history and knowledge documents have different trust levels. User turns can contain mistakes, prompt injection, or personal data; retrieved documents should carry source metadata and citation rules. I would also evaluate it with multi-turn cases: pronoun references, changed user intent, contradictory prior turns, and questions where history should not influence retrieval.