AI, Agents, and Spatial Presence
Telemus AI combines a workspace assistant with helpers for research, review, planning, and other tasks. Supported conversations can use visible documents, project context, notes, and workspace history. See Product Essence.
Your workspace assistant
Telemus is the assistant. Ember is its default appearance. Welcome and
Settings also offer Azure, Lumen, Moxie, Sprig, Sunny, Sol, Echo, and Pip.
These names select appearance, not an AI provider. Native Welcome shows a model
preview; Web shows thumbnails; the terminal uses /avatar and /avatar <id>.
The saved preference is shared across those surfaces.
The packaged dot characters have static faces and whole-body motion. They do not provide articulated walking or facial lip-sync. Other avatar capabilities depend on the selected model and installed release. If a character cannot load, the app reports its fallback or unavailable state. Custom character upload is not available from the picker.
You can ask the assistant to:
- summarize a supported document or extract action items;
- create notes, update a task board, or organize a project;
- identify a visible document from a question such as “What is in that PDF?”;
- find related workspace material and bring it into view;
- speak a response using your configured voice provider.
Speech, spatial audio, and avatar motion vary by device, model and provider. Some system voices play without positional audio or mouth animation. Check voice status rather than assuming every provider has the same capabilities.
Helper agents
Create a helper for a bounded task and inspect its progress while you continue working. Helpers can appear in the room with their own status and transcript. Supported helpers can move toward work surfaces, use assigned zones and return results to notes, task boards, or other artifacts.
Telemus Operations shows channel sessions and subagents. Agent Sessions provides conversation inspection. A helper reporting completion does not independently prove that its answer or external action is correct; review the result and any accompanying evidence.
Use spawn_pool_status to inspect available helper capacity, agent_hierarchy
to inspect supported parent/child relationships, and agent_roles to inspect
available role names. A role name does not grant permissions.
Supported Hermes integrations offer Load Hermes and Switch to Hermes.
Terminal users can inspect the configured integration with /hermes and
/hermes-load. These require the corresponding service and session to be
available. An external agent's own execution and cancellation rules may differ.
Reusable skills and memory
Skills are reusable procedures. The assistant can discover a relevant skill,
open its instructions with skill_view, and use available tools to carry out
the work. Installing a skill does not authorize every action described in it.
Only install skills and executable support files you trust. See
Skill Library.
Supported memory tools include memory_search, memory_get, remember_this,
forget_this, and session_search. sessions_list, sessions_history, and
delegate_subagent are also available in supported messaging-channel workflows.
Availability and access remain subject to the current account, host and policy.
user_model_status is not available to messaging channels.
Local memory is the default. Configured external providers such as Honcho,
mem0, or Supermemory can receive recall requests. Operators can require
confirmation for external memory reads with action_policy.memory_read_ask=on.
Choosing local memory alone does not guarantee that every AI operation is offline.
forget_this distinguishes removing a preference from erasing stored content.
A successful preference removal is not a promise that every copy was erased.
Review the reported result and any storage errors.
Workspace instructions
Trusted local workflows can use workspace-owned SOUL.md, IDENTITY.md,
AGENTS.md, USER.md, MEMORY.md, and CHANNEL.md. These describe identity,
preferences and workspace conventions. TOOLS.md and BOOTSTRAP.md are optional.
Use reload_persona after editing supported instruction files. Private workspace
instructions are not automatically shared with messaging-channel users.
Keep credentials out of instructions and memory. Saved instructions and skills do not replace tool permissions, action approvals or your current request.
Reading your own sessions and files
Where host file access is enabled, Telemus can inspect supported agent sessions and files available to the account running the app. An explicit workspace scope restricts access to that workspace. Without one, supported text/code reads may reach other files the operating-system account can read. PDF access has its own workspace and account-home limits.
Known credential locations are excluded, but a path check is not a content security guarantee. Do not assume that a renamed credential file is protected. Operating-system permissions remain important. Session searches are bounded; truncation or unread transcripts are reported rather than presented as a complete account history. Observing another agent session does not grant control of that session.
Long-running work and interruptions
Long work remains subject to configured time, tool, context, provider and concurrency limits. When work stops at a limit, inspect the progress and use continue when appropriate. This resumes from available context; it does not promise exact replay or uninterrupted execution.
On supported local channel hosts, a long shell command may return a process
handle so the assistant can check its existing work. An explicitly detached
background command can have a different lifetime from the conversation. Check
its status before starting another copy. /stop and /new stop or clear work
according to the current host's cancellation support; they cannot undo effects
already completed or guarantee that a process died after an app crash.
Messaging-channel conversations retain bounded recent context. Older details
may be summarized or omitted. /new clears the channel's local conversation
context. This is not a full archive or a promise of perfect recall. Direct Web,
terminal and native conversations can have different history limits.
If a provider fails or returns no usable answer, the app reports a failure or available progress. Before retrying, inspect any tool changes already made. Choose a model with adequate context and output capacity for the task; the app can reject models whose context window is too small.
Review and control
Actions remain subject to the configured approval policy. Supported output can include evidence and an audit trail. A prepared artifact is not an authorization to send it, and a status message is not proof of successful delivery.
For supported ACP integrations, Telemus can send a structured request to a configured ACP-capable process. Interactive terminal programs require a compatible integration; an ordinary terminal program is not automatically an ACP agent.