1Chat is built around a simple promise: your AI workspace should carry useful context between models without turning that context into a black box. This release tightens the parts of the product where that promise matters most — what a response used, what memory is allowed to influence, and whether the same workspace is usable away from a desktop.
A Context panel that reports what actually happened
The response Context panel now comes from the turn's execution record. It clearly separates four different things:
Context injected — memories, skills, knowledge, and files placed into the prompt.
Tools called — the actions that actually ran, including whether each succeeded.
Tools available — capabilities the model could have chosen but did not necessarily use.
Sources returned — links produced by searches during the turn.
That distinction matters. Turning web access on does not mean web search ran. Having a memory tool available does not mean it was called. A memory search can no longer be reported as "no memory used," and an enabled toggle can no longer masquerade as execution.
Context injected2 memories · product brief.pdf
Tools calledsearch_memory · succeeded
Tools available12 connected actions
Sources returned3 links
The Context panel now reports what happened, not what was merely switched on.
Safer, cleaner memory
Learned lessons can improve tone and workflow, but they are the wrong place to establish security policy. Statements that resemble authentication, permissions, approval rules, secrets, or attempts to bypass confirmation are now placed behind a protected policy boundary. You can inspect them in Memory, but 1Chat will not inject them into model behaviour.
The memory facts view also does more consolidation before showing graph output. Exact duplicates are merged, repeated words are cleaned up, related-source coverage produces a visible heuristic confidence level, and singular claims such as a workplace or favourite with different values are marked as conflicts. The goal is a memory surface you can correct, not a raw graph dump you have to decipher.
Every Space works on mobile
The mobile drawer now exposes the actual Space architecture: switch Spaces, start a chat, and reach Knowledge, Memory, Tools & Skills, and Space settings. Model controls wrap instead of disappearing off-screen, while advanced controls move into the existing mobile control sheet.
The desktop remains the best surface for dense work, but mobile no longer hides the features that make a Space a Space.
Less catalogue, more choice
The Models page now opens on a short Recommended view built from current Anthropic, OpenAI, and Google models. The complete catalogue and raw model IDs remain available under All models for people who want them.
Tools & Skills also gained a direct add-tools flow for each Space. You can attach an account connection you already have or search Composio and connect a new app without leaving the Space. Imports from existing chats bring proposed memories, reusable skills, and relevant connection suggestions into the same review flow before anything is applied.
Better feedback when work is still happening
Loading states now preserve the shape of the page instead of presenting an unexplained blank screen. Empty states point toward the next useful action. Deliberation steps that were intentionally skipped are neutral rather than red; red remains reserved for a real failure.
This release deliberately adds very little breadth. It makes the existing system easier to understand, safer to trust, and more complete wherever you use it. That is the standard we want for the releases that follow.
Try it yourself
Every model, one memory, and agents that work the way you do.