What the AI Assistant Remembers After the Chat Is Closed
People expect an assistant to remember the relationship. This one remembers less than they think, and that is deliberate. Firmspace AI has two separate memories: a short conversation window that disappears, and a small notebook of rules the user taught it. Everything else — customers, invoices, deals — is read from the CRM again on each request.
The web assistant keeps its messages in the open tab only. The conversation is not saved on the server and not saved in the browser; reloading the page or closing the tab ends it. The browser sends the last twenty messages with each request, and the server keeps the last twenty; empty entries fall away. The reset button clears the thread on purpose.
Telegram works differently because the server polls the bot. The Iris bot keeps a conversation in process memory for thirty minutes after the last message. The model receives the last eight earlier messages plus the current request; the pending confirmation lives in the same state but is not part of the transcript. A restart of the server clears that state, and /reset clears it earlier. If the bot was offline for a while, messages older than ten minutes are ignored rather than executed late.
On every turn the model sees the current request, the recent messages that survived the limits, the page the user is looking at, the user and firm name, the local date and time, and the saved memory block. It does not see yesterday's chat, the whole mailbox or every deal. When a fact is needed, the assistant must call a tool and read it; that read is logged.
The long-term memory is not automatic learning. It is a small notebook the user teaches on purpose. When someone explains how they name things or corrects the assistant, the instruction says to save that as a rule immediately and silently, one short self-contained rule per entry, up to four hundred characters. Nothing is copied from the conversation in the background.
The notebook has hard limits. One user can hold up to eighty entries; duplicates are not saved; when it is full, the oldest entry makes room for the new one. In the prompt the newest rules are kept first until three thousand characters are reached, and the block is then shown oldest to newest. A very old note can stay in the database and still not reach the model. The limit is explicit, which is better than a memory that silently grows.
Memory belongs to a person, not to a firm. It is stored per user inside the company, so a rule learned in the web app follows that person into Telegram and back. A colleague's assistant does not inherit it, and a new employee starts with an empty notebook. That is useful for individual style; team standards still belong in written instructions, not in one person's memory.
What belongs in the notebook is how work should be done: that CRM means the deals board rather than the dashboard, which language to answer in, how to format a price, which report is wanted on Monday, how the firm names its stages. What does not belong there is what is true about a customer: balances, promises, delivery dates, who pays late. Those are read from the card and the invoices each time, because they change.
Memory can go stale, and the assistant is told to prefer current data. If a note says a customer always pays late but this month's invoice was paid on time, the invoice wins. A new instruction is supposed to beat an old memory: the assistant is told to forget the outdated rule and save the new one. That is a rule in the prompt, not a guarantee; when a note is wrong, say so and ask for it to be removed.
The notebook is visible and editable. Ask “what do you remember” and the assistant lists the entries. Remove one with a few words from it or with its id; a phrase that matches several entries removes all of them. This is not a hidden profile built from behaviour — it is a list the user can inspect and correct.
Two actions are easy to confuse. Reset clears the current conversation; it does not touch the notebook. Forget removes long-term rules; it does not clear the thread. If an assistant seems to remember something after a reset, the note is probably in the notebook, and the right tool is forget, not reset.
Privacy has a clear shape. The chat text is not archived by the assistant, but tool calls are logged for audit: what was read, proposed, executed or failed, with arguments and a short result. The notebook is stored as text in the firm's CRM database, not in a separate secret vault. Do not teach it passwords, card numbers or access tokens; it is a place for working rules, not for secrets.
A practical habit: reset before a new topic, correct in one sentence when the assistant gets something wrong, and put customer facts in the card rather than in the chat. Use the notebook for how you want work done; use the CRM for what is true about a customer. The two memories then do different jobs instead of fighting each other.
Small memory is a feature, not a defect. The assistant is useful because it reads the current CRM and follows the rules it was taught, not because it pretends to remember every conversation. When it says it does not know, that is often the most reliable sentence in the chat.