Memory across conversations: how many to recall and max age

Make your agent remember a customer's past chats: how many conversations to summarise and how far back to consider them.

Jun 26, 2026

Watch this step in the video · Memory across conversations and summary on close

What it is and when to use it

Memory across conversations lets your AI agent remember what happened in previous chats with the same customer. When that person writes again days or weeks later, the bot already "knows" what was discussed, what they bought or what issue they had, without the customer having to repeat it.

It works through summaries: every time a conversation is resolved (closed), Omnifox automatically generates a summary of what happened. When a new conversation from the same contact starts, the agent receives those summaries in its context.

Important: each close generates a summary, which costs 1 AI turn. It's a small cost, but worth keeping in mind at high volume.

How to set it up

  1. From the side menu, open AI Agents and select the agent.
  2. Go to the Behavior tab and scroll to Memory across conversations.
  3. Turn on Enable memory of previous conversations.
  4. Under How many prior conversations to summarise, set how many past chats the bot considers (between 1 and 10). More conversations = more context, but also more text in the prompt.
  5. Under Only conversations from the last N days, set the maximum age (between 1 and 365 days). Older conversations are ignored even if they exist.
  6. Save the agent.

How to pick the values

  • How many (1-10): for support, 3-5 is usually enough. For consultative sales with long history, raise it.
  • Age (1-365 days): if your product changes fast, lower it so it doesn't drag stale data; for long-term relationships, raise it.
  • The bot uses the most recent ones within that limit, so both controls work together: "the last N conversations, as long as they're from the last M days".

How to verify

  • Resolve (close) a conversation with a contact so its summary is generated.
  • Start a new conversation with the same contact and, in the sandbox tester or a real chat, check that the agent references the earlier interaction.

If it doesn't work

  • The bot remembers nothing: confirm the earlier conversation was closed/resolved (summaries are generated on close) and that it falls within the configured day window.
  • It brings up very old data: lower Only conversations from the last N days.
  • It mixes customers: memory is always per same contact; make sure you aren't testing with different contacts.
  • AI usage goes up: remember each close generates a summary; see the cost article.

Tip: start with 3 conversations / 90 days (the defaults) and adjust as you see response quality.

Tip: pair this memory with solid agent instructions so it knows how to use the history (greet by name, resume the pending order, etc.).

Related articles

Was this helpful?

Related articles

Message us on WhatsApp