I
Ignacio Montenegro
Summary:
Membase currently treats connecting or importing a source as a high-risk, hard-to-reverse action. There is no way to filter what enters the system before processing, no efficient way to clean up after a bad import, and no control over what gets inferred from a connected source. These four requests address the full data lifecycle and together form a memory control layer that would make importing feel safe, trustworthy, and empowering.
---
1. Pre-Import Filtering & Curation
When importing Claude or ChatGPT conversation histories, there is no way to review or filter content before it is processed. For accounts shared between multiple people, this means Membase incorrectly attributes memories from other users to the account owner — polluting the profile from the start. Beyond shared accounts, most users would not want to import everything blindly. Irrelevant conversations, sensitive topics, and outdated exchanges all introduce noise that degrades memory quality.
Proposed solutions:
  • A pre-import conversation where the user can give context before processing (e.g. "conversations from this person are not mine — ignore them")
  • Inline annotation to flag or exclude specific conversations or date ranges before ingestion
  • A review UI to skim what would be imported and select what to include or skip
---
2. Bulk Deletion After Import
After connecting a calendar, the resulting memory profile contained significant noise — one-off events, past appointments, spam invites, and irrelevant entries. There is no efficient way to clean this up. Deleting items one by one is impractical at scale. Without bulk delete, users who know a messy import is hard to undo will hesitate to connect high-volume sources at all, which directly reduces feature adoption.
Proposed solutions:
  • Filter memories by source, date range, or type — then bulk-select and delete
  • A "delete everything from this source" action for clean removal of a specific import
---
3. Extraction Intent When Connecting a Source
When connecting a calendar, Membase extracts everything indiscriminately. But a calendar contains many different types of signal — social connections (who you meet and how often), behavioral patterns, professional context, and a large amount of noise. Different users want different things from the same source. This applies equally to Gmail, Slack, Drive, and any high-volume source.
Proposed solution: a configuration step when connecting any source, where the user chooses what Membase should learn from it — high-level patterns and relationships only, full detail, social graph only, or a custom scope. People are handing over sensitive data, and giving them a say in what gets inferred builds trust. Connecting a source should not feel all-or-nothing.
---
4. Post-Import Attribution & One-Click Account Splitting
As a fallback to pre-import filtering, it would be valuable to support post-import attribution — the ability to assign imported conversations to a specific person, so Membase correctly routes memories and wiki entries to the right profile. If two people shared an account, they could tag each conversation after import: "this one is mine, this one is yours."
Taking this further: a one-click account split feature, where each person's attributed memories and wikis are exported into a new, separate Membase account.
**Hard blocker worth noting: Claude and ChatGPT only support one connected Membase account per AI account. This means that if two people share an AI account, neither of them can use Membase properly. The person who connects their Membase account ends up with a contaminated memory profile — their conversations are mixed with the other person's, so Membase builds the wrong context and starts giving inaccurate, misleading replies. The other person cannot connect Membase at all. So it is not a situation where one person wins and one loses — both users get a broken or nonexistent experience. These are not edge case users. They are people who are fully blocked from getting value out of Membase today.
---
Business Perspective
At first glance, supporting shared AI accounts might seem to reduce paying customers. The logic actually runs the opposite way. People sharing accounts on ChatGPT or Claude will not adopt Membase if it cannot handle their reality. If Membase solves this well, it converts one shared AI account into two or three individual subscriptions, as each person discovers the value of their own clean memory profile. The account split feature makes this conversion explicit and frictionless. Shared AI accounts are more common than they might appear — families, couples, and small teams routinely share access. This is not a niche edge case. It is a currently unreachable segment that these features would directly unlock.