AutomationAI Colleagues

Colleague Conversation Memory

Your AI Colleagues keep a rolling memory of each guest conversation, so they pick up where they left off instead of re-reading the whole transcript on every reply. This article explains what the memory contains, where to see it, and how to refresh it.

What changed

Previously, every time a guest sent a message, the colleague re-read the entire conversation history before composing a reply. That worked for short threads but caused two problems on long ones: replies got slower, and the colleague sometimes treated its own prior messages as authoritative facts, repeating itself or ignoring newer information.

Now the colleague keeps a short structured summary of the conversation — guest intent, established facts, open questions — and replies using that summary plus the most recent few messages. The effect is faster, more focused replies that don't lose track of what was already established.

What the memory contains

The memory is a structured snapshot of the conversation. Four fields:

  • Intent — one short phrase describing what the guest is currently trying to do (e.g. "asking about late check-in").
  • Summary — three to five sentences a human can read to catch up on the thread.
  • Established facts — short list of things the guest or your team explicitly stated. Capped at eight items so the memory stays tight.
  • Open questions — questions that have been asked but not yet answered.

The snapshot updates automatically a few seconds after each outbound reply. The colleague uses the latest snapshot the next time the guest writes.

Where to see it

Every conversation in the Inbox has a Conversation memory card pinned to the top of the right sidebar. It shows the four fields above plus a last updated timestamp.

The card is most useful when:

  • You take over a thread mid-conversation and want to catch up without scrolling through every message.
  • You want to verify the colleague has understood the guest correctly before letting it reply again.
  • You're reviewing an escalated case — the summary doubles as the operator brief.

Refreshing the memory manually

The memory updates automatically after each colleague reply, but you can also recompute it on demand using the small refresh icon in the top-right corner of the Conversation memory card. Use this when:

  • You added context to the conversation by replying as a human and want the memory to reflect what you just wrote.
  • The memory looks out of date because the automatic refresh hadn't fired yet.
  • You're checking a thread that has never been replied to by a colleague — the memory will be empty until you trigger the first refresh.
The refresh runs in the background and takes one to two seconds. The card will show a small spinner while it's working.

Syncing a conversation from Hostaway

Separately from the conversation memory, the inbox thread now has a Sync button in the conversation header. This re-pulls the messages from Hostaway and updates the local copy.

Use it when:

  • A guest message seems to be missing from the inbox but is visible in Hostaway.
  • A reply you sent from Hostaway directly doesn't show up here.
  • You suspect the live webhook missed an event (rare, but possible during Hostaway outages).
Most of the time the webhook-driven sync runs automatically and you don't need to press Sync. It's a fallback for the rare cases when the automatic path missed something.

What this changes for your colleague setup

Nothing in your colleague configuration needs to change. The memory is built and consumed automatically — your existing details, knowledge base access, and filters continue to work exactly as before.

One practical effect to be aware of: your colleague's replies in long threads will become noticeably more focused on the current question and less prone to repeating earlier answers.

Limits and edge cases

  • The first reply on every conversation still uses the full transcript — the memory only exists after the first colleague turn. You'll see meaningful values in the card starting on the second inbound message.
  • If the memory compression fails for any reason, the colleague falls back to using the full transcript for that turn. No outage, just a degraded mode you won't notice.
  • The memory only stores what was actually stated in the conversation. It will not invent guest names, dates, or amounts that weren't mentioned.