# STEM: selected context across assistants

Silent captioned illustrated replay, 42 seconds. Based on one isolated synthetic test observed October 25, 2026. This is not a screen recording. The fictional illustration is not the original test fixture.

## 00:00:00–00:00:05

SELECTED CONTEXT, ACROSS ASSISTANTS

Keep the context.
Change the assistant.

STEM by Awase

ChatGPT  →  Claude  →  Cursor

One authorized STEM account.
A selected note, recalled elsewhere.

Illustrated replay • Based on one isolated synthetic test • October 25, 2026

## 00:00:05–00:00:11

01 / CONNECT

Connect each assistant
to the same account.

Complete OAuth and review permissions.
For this test: only save when asked.

YOU CHOOSE WHAT TO SAVE

Approve a useful note.
Keep control of what is shared.

Selected context only. This is not full-chat import or live transcript sync.

## 00:00:11–00:00:18

02 / SAVE IN CHATGPT

Save the decision
you want to carry forward.

ChatGPT reported a successful save.
Its underlying save response was not
independently captured.

FICTIONAL ILLUSTRATION

Project Juniper:
Use a simple onboarding flow.

Illustration text differs from the private test note. No actual chat UI is shown.

## 00:00:18–00:00:25

03 / RECALL IN CLAUDE

Start a fresh chat.
Ask for the saved context.

Claude’s native tool response returned
the exact expected test note and
the reported memory ID.

OBSERVED NATIVE RECALL

The receiving prompt supplied
only a lookup label —
not the saved answer.

One synthetic test on October 25, 2026. This is an illustrated replay.

## 00:00:25–00:00:32

04 / RECALL IN CURSOR

Bring that selected
context into your work.

Cursor’s native recall also matched
the expected note and reported ID.
It used the User MCP connection.

SCOPE OF THIS TEST

OAuth-authorized recall verified.
Cursor saving and installed-plugin
runtime were not tested.

One observed run; not a reliability or latency benchmark.

## 00:00:32–00:00:37

05 / REVIEW BEFORE USE

Recall. Review.
Then continue.

Check the saved context is still relevant.
Ask before saving new information.

CLEAR BOUNDARIES

Selected notes across assistants.
No full-chat import.
No automatic transcript sync.

The demo used an isolated account with synthetic information.

## 00:00:37–00:00:42

START WITH THE AGENT HUB

Connect your assistants.
Keep useful context.

Setup guides, capabilities and
the dated evidence behind this demo.

STEM BY AWASE

awase.dev/agents

Connect with OAuth.
Review permissions. Choose a note.

Silent, captioned explainer • Illustrated replay, not a screen recording

## Evidence and limits

Evidence source: the sanitized isolated native-assistant receipt and report in this repository. The ChatGPT save acknowledgment is assistant-reported. Claude and Cursor native recall responses independently matched the expected note and reported memory ID. Cursor used its User MCP configuration; saving through Cursor and installed-plugin runtime were not tested. One synthetic run is not a reliability benchmark. No full conversation import, provider-native memory transfer or live transcript synchronization was established.

Public setup and evidence: https://awase.dev/agents and https://awase.dev/workflows/chatgpt-to-claude