Conversations with context.

Continuity should be useful, not mysterious

A chat agent neural API connects language responses to conversation context. The difficult product question is what that context should contain. A current request, a confirmed preference, an old suggestion, and a retrieved document are different kinds of information. Treating them as interchangeable can make a coherent-sounding answer rely on an outdated assumption.

Design the memory policy before adding a permanent store. A session-scoped support assistant can be useful without retaining a profile of everything a person has said. Add durable information only for a clear purpose, with an understandable way to inspect and correct it.

Keep three context layers distinct

The active conversation

Preserve the user's current goal, confirmed constraints, attempted steps, and unresolved questions. When compressing a conversation, retain corrections and rejected suggestions rather than summarizing only the assistant's last answer. Keep pointers to important source messages where the implementation permits it.

Durable settings

Separate an explicitly confirmed account preference from an inference. Record purpose, provenance, and scope. A temporary delivery address or a request for a shorter answer today should not silently become a permanent setting. Give the user a clear way to see which assumption is influencing the response.

Retrieved knowledge

Bring in authorized documents and records only when they help the current task. Preserve version and source information. Check access before returning material to the model; do not retrieve broadly and rely on the assistant to hide protected details afterward.

Make corrections stronger than old context

When a person corrects a product model, mark the previous value as superseded. Repeated older mentions should not outweigh a newer explicit correction. Define the scope of the update so a session instruction does not overwrite an unrelated account setting.

Test a long conversation with a crucial correction early in the history. Add irrelevant documents and verify that the assistant still respects the current task. Evaluate these failures separately from general writing quality: a well-worded answer based on the wrong device is still a failed support interaction.

Design a clean handoff

A person taking over should see the confirmed goal, relevant evidence, attempted steps, and unresolved question. Mark a generated summary as a summary, not a verbatim record. When history is missing, ask the user for the necessary context rather than inventing continuity.

Keep retention and deletion decisions consistent across conversation records, summaries, logs, caches, and retrieval indexes. A delete control should describe what it actually changes. The examples here create no chat history and send no messages to an AI provider; they show how to reason about a real implementation before collecting user data.