Assistant Memory Summarizer: keeps long chats useful
Long conversations get heavy. After twenty messages, the early details are buried, and sending everything again every time makes the assistant slower and less focused. Assistant Memory Summarizer solves that. It condenses an assistant conversation into a short note, and the next turn starts from that note plus the latest messages.
You do not run it yourself. It works in the background of the page assistants, which is how Worknano stays focused on core functionality while AI quietly handles the housekeeping. The assistants behave the same across modules, with the same UI, UX and database underneath.
What it does
- Takes the existing summary and the recent turns of a conversation and returns one compact summary.
- Keeps what matters for the next step: your preferences, exact model and record names or IDs, confirmed facts, unresolved questions and requested actions that are still pending.
- Drops what does not: greetings, repetition, raw table rows and details that no longer apply.
- Stays under about 1,200 characters, so the note is cheap to carry along.
- Does not add facts that were not in the conversation.
When it helps
- A long planning session. You have spent an hour with the OKR page assistant, refining key results. When you ask for the next change, the assistant still knows which objective you mean.
- Coming back later. You return to a conversation the next day. The assistant starts from the compact note instead of the full transcript.
- Pending actions. You asked for three updates and confirmed one. The note keeps the other two, so they are not lost.
How it works with your data
Assistant Memory Summarizer has no tools. It does not read your records and it does not write to them. It only sees the conversation text it is given, and it returns text. The assistant that called it uses that text as memory at the start of the next run. Anything the assistant reads or changes still goes through the signed-in user's permissions and, for writes, through a preview you confirm. The summarizer cannot give an assistant access to anything the user could not see.
It runs on a fast model, because compacting text is a small task.
What it does not do
- It is not a record of truth. The conversation is, and the summary is a shortened reading of it.
- It does not decide what you did. Records stay as they are in your data, not in the note.
- It does not keep everything. Raw tables are removed on purpose, so ask again if you need the rows.
It works hand in hand with Generic Assistant Planner and the specialised page assistants such as Task Assistant.
Make it yours
Assistant Memory Summarizer is a record in the Agents module. If your team always needs certain details kept, such as ticket numbers, add that to its instructions. You can also change the model, duplicate it, or start from a blank agent.
Start with Build your own agent, read How to edit a page assistant and a reusable agent for the editor steps, or see the bigger picture in AI agents and AI assistants.