Skip to content

Review Bot memory

Bot memory keeps facts and observations from earlier work. It is separate from the message history of a Chat. Learning needs memory to be enabled and a memory model to be set by your administrator.

Learning has its own durable execution. It waits for the Turn that asked for it. A restart keeps that work. The saved execution identity sets who may see each fact. Memory text, confessions, and input additions are encrypted in storage.

All server instances share the memory model limit for a Bot. A new request counts work that is already running, even after the limit changes. Waiting is part of the time limit: two minutes for learning and 15 seconds for other memory requests. A restart waits for the old request to release its place or reach its saved deadline. Closing a request cannot prove that the provider stopped its work. Usage stays unknown until the provider confirms it.

Facts record what work showed. Observations are conclusions from other entries. Pinned entries are things a person asked to keep. An observation can be wrong; check it against its sources.

Use the memory functions of your client to read entries you may see and change those you may manage. Correct old facts and remove entries that should no longer be kept. Deleting Chat history does not remove these entries.

An entry has an audience. Some are for one person, some for a Channel group, and some for everyone using the Bot. Group work can use only memory fit for that group. Sharing a Bot does not let its owner read private memory about another member.

You may read a conclusion only if you may read every source that supports it. Taking part in one source is not enough. Old memory whose source access cannot be checked stays hidden until it is rebuilt from checked sources. Ask your administrator if learning is unavailable; ask your client provider if memory controls are missing.