Conversation context and memory

Pneum.ai uses several different forms of context. Knowing which one you are changing prevents accidental loss and avoids assuming the assistant remembers something it does not.

Conversation context

The active transcript supplies recent requests, responses, tool results, and coordination events to the model. Long conversations consume more of the model’s context window.

The context indicator reports usage. When configured thresholds are reached, the agent can compact older turns into a summary while retaining recent messages. Review the compaction dialog when it is shown; a summary preserves intent and decisions, not every verbatim detail.

Profiles are behavior, not separate disks

Profiles change identity, model, speech, and activation settings. They do not create a new cloud computer. Use separate Cloud agents when you require separate files, credentials, integrations, and runtime state.

Notes are durable knowledge

Store facts, plans, and reference material in Notes Studio when they should outlive the immediate conversation. Markdown notes can be searched and intentionally brought back into a request. Do not rely on a very old transcript turn as the only copy of an important decision.

See Notes Studio.

Vault stores generated media

Vault contains generated images, audio, music, and video metadata and files. It is not a transcript memory system. A Vault asset can be downloaded or used as chat input without making every future conversation aware of it.

Channel isolation

Supported external chat connectors create source-scoped conversations. Direct messages are isolated by immutable platform identity. Group chats, channels, and threads use their own scoped contexts. The private owner conversation is not copied into an external participant’s context.

The owner can open an admitted channel conversation and steer within that context. Quoted channel content remains an untrusted observation rather than an administrator instruction.

Reset carefully

Starting a new conversation clears or separates model context; it does not automatically delete notes, Vault assets, connector credentials, or the Cloud VM. Termination is the operation that deletes the agent computer and its local data.