Memory and context management

What your agent remembers and works from each turn — conversation history, long-term memory per user, knowledge retrieval, and automatic handling of very long conversations.

Each turn, your agent works from several sources: the current conversation, its long-term memory of the user, and your knowledge base. This page describes what your agent remembers, how it’s kept separate per user, and how long conversations are handled.

What your agent works from

Source	Lifetime	What it is
Conversation history	The current conversation	The back-and-forth so far
Long-term memory	Until erased	What the agent has learned about this user
Knowledge retrieval	Until re-ingested	Your knowledge base, searched each turn

Every conversation is retained in full, so nothing is lost even when a channel only carries recent messages on the wire. Instant-messaging threads carry their full history; web and widget chats carry a recent window and rely on the retained transcript for the rest.

Long-term memory

Your agent builds up memory about each user over time. The controls live on the Memory page, and the full guide is User memory. In brief:

• Two kinds of memory — freeform notes about what happened, and a structured profile of durable facts.
• Confidence. Memories carry a confidence level. Things the user stated directly count for more than things the agent merely inferred, and low-confidence, stale notes fall away over time.
• Conflicts are flagged. When a new memory contradicts an existing one, the agent surfaces it with a caveat instead of asserting it, until you resolve it from the memory page.
• Learned automatically. After a conversation, the agent quietly distills what’s worth remembering, without slowing down the reply.
• Separated per user. Each way a user reaches your agent — dashboard login, API key, IM sender, email sender, contact — has its own private memory. What the agent learns about one user never bleeds into another.
• Owners can’t read a subscriber’s contacts’ memory. When a subscriber runs their own public instance, their visitors’ memories are kept private from the owner — a real privacy boundary, not just a hidden view.

Knowledge retrieval

Each turn, your agent searches your knowledge base for relevant passages and uses the best matches in its answer. You control how strict retrieval is — a relevance floor and optional reranking — from the knowledge controls, and you can see exactly what was searched and used in each turn’s activity detail. To try a query interactively, use the Retrieval Test page.

Long conversations keep working

A very long conversation doesn’t fail — as a thread grows, your agent automatically condenses the older part into a summary and keeps the recent messages in full, preserving the important details (the user’s preferences and constraints, decisions, open tasks, and concrete facts). This happens on its own, and any cost is shown in that turn’s cost breakdown.

If a single message is genuinely too large to process, your agent returns a clear “input too large” error rather than silently dropping content.

Memory is screened

Because remembered facts come back on every future turn, anything your agent is about to remember is screened for hidden instructions first. A flagged item is dropped rather than stored, so a poisoned document can’t plant an instruction that re-injects itself later. See Safety and isolation.