The Center for Democracy & Technology released Selective Memory on 28 October to address a practical problem with changing chatbots: a conversation archive can contain useful project history alongside information its owner wants to leave behind.[1]
The prototype lets users select conversations, remove messages and redact passages before generating a new export. CDT presents it as a demonstration of controls that AI providers could build into their own import and export processes.[1]
A local editing step
The documented application runs on the user’s computer and accepts conversation exports from Claude and ChatGPT. After selecting the source platform and loading a conversations.json file, users can search and sort the chats, edit their contents and download the chosen subset.[2]
What the sensitivity colours mean
| Label | Rule-based indication |
|---|---|
| Red | Potentially high-risk content, including health, finance or live secrets |
| Yellow | Potentially identifying or commercially sensitive content |
| Green | No configured pattern matched |
A green label therefore means only that the rules found no match. The documented workflow includes human review and manual redaction of additional material.[2]
Editing an archive and importing it are separate steps: the receiving service still needs to accept the file. CDT also sees a use for selective exports in research, allowing volunteers to choose which conversations they donate and remove details they do not want to share.[1]
Sources & further reading
- Selective Memory and AI data portability Center for Democracy & Technology · 28 Oct 2026 · Accessed 28 Oct 2026
- Selective Memory documentation CDT AI Governance Lab / GitHub · Accessed 28 Oct 2026