In one sentence
The memory of the AI legal chat is off by default, every entry is visible and can be deleted individually — and the management view tells you openly that only the 30 most recent entries reach the model at all.
The default setting is the real decision
With a memory for an assistant there is exactly one place where you have to be honest: the default setting. Everything else — settings page, delete button, explanatory text — comes afterwards and changes nothing for the people who never go into the settings.
Our default is off. The reasoning is written into the code just like that: the platform learns nothing across a person until that person actively switches it on. Anyone who uses the chat and never touches the function leaves no profile behind — not because a tick box could be clicked away, but because nothing starts up.
What is learned once you switch it on
Even switched on, the learning stays deliberately sparing:
- At most four entries per run.
- Only where confidence is high or medium — anything uncertain is discarded rather than kept.
- Duplicates are skipped.
- The run starts only from the second message onwards, and after that applies to every third one.
These caps are not thrift for its own sake. A memory that records everything quickly becomes a collection of half-true sentences about a person — and the larger it grows, the less anyone notices what is in it.
Reading and deleting — but not writing
A detail that looks like a restriction and is in truth a safeguard: you may read and delete your entries, but not create them yourself. Only the server is allowed to write.
The reason lies on the attack side. A memory a client can write into is a memory into which something can be written — through a prepared page, a manipulated script, a slipped-in document. What the model later treats as a known preference would then no longer be what you said. That is why the way inwards is closed.
The limit of 30 is stated in the interface
At most 30 active entries go into the prompt, the most recent first; individual values are shortened to 300 characters in the process.
We could have kept this figure quiet — the management view could simply have shown 30 entries, and nobody would have asked. Instead it shows all entries and marks the limit visibly: from position 31 onwards the row reads “no longer in the AI context”. An entry you see there exists — it merely no longer influences the answer. That is more uncomfortable and more correct than a list that pretends everything is effective.
Who else reads along
Two further functions access the memory: the generation of letters, and document analysis. Both only read it. Writing happens exclusively from within the chat. There is thus exactly one place where entries arise — and exactly one you have to watch if you want to know where something comes from.
What the memory is not
- It is not case knowledge. What is in your documents is a different level; what lies here are preferences and recurring context, not facts of the matter.
- It decides nothing. A remembered circumstance changes the cut of an answer, not its legal substance.
- It does not make answers more binding. Even with full context the chat remains an orientation aid; where the individual case matters, the way leads to a law firm.
Anyone who does not want the function need do nothing. That is the point.
→ Free AI legal information · AI preferences · Security · AI transparency
