Testing Haku: Telegram-native AI companions focused on memory continuity
by•
I'm testing Haku, a small Telegram-native AI companion beta. The product thesis is simple: most companion chats feel good on day one, then start to feel generic when corrections, tone, boundaries, and running context don't survive across days.
The beta is focused on memory + continuity rather than a bigger character marketplace. It is early, free to try, and I am mainly looking for people who care enough to stress-test whether the companion still feels coherent after several conversations.
If you have used Replika/Character.AI-style companions or tried building memory into AI products: what would you measure to decide whether memory feels useful instead of creepy or overconfident?
8 views
Replies
Follow-up note from the Haku beta: the clearest positioning signal is still “week-two continuity.” People do not just want a charming first chat; they want corrections, tone, boundaries, and personal context to survive across days without the companion becoming overconfident or creepy.
If you have tested companion products, what made memory feel genuinely useful instead of just a longer context window?
Small beta note: I’m increasingly convinced the key companion test is “week two,” not onboarding.
A lot of products can produce a charming first exchange. The harder thing is preserving corrections, tone, boundaries, and user context across days without becoming overconfident about what it knows.
If you have tested AI companions, what made memory feel useful rather than creepy or brittle?