I gave my iOS reply coach a memory (and what broke)

Description and use

Toneme added contextual memory by storing short notes about recipients, improving reply relevance, but faced model hallucinations, relayer rate limits, and duplicate notes.

Problem

The reply coach started each message from zero, lacking context about the recipient and prior interactions.

Outcome

The app gained contextual awareness, improving reply relevance, but encountered issues such as model‑generated meeting times, relayer rate limits, and duplicate notes.

Workflow

• Add opt‑in memory using Walrus Memory
• After each reply, save a short note (e.g., "Nikos is accounting colleague; user writes briefly")
• Recall up to five notes before the next reply
• Provide UI to view, edit, or reset notes
• Store each note as an encrypted blob on Walrus mainnet and link to Walruscan for inspection

Evidence

Its problem: every message started from zero. It never knew that Nikos is a colleague or that I write to him briefly.

Source

After each reply, the app saves one short, non-sensitive note ... and recalls up to 5 notes before the next reply.

Source

The model invented a meeting time.

Source

The shared relayer rate-limited us.

Source

Near-duplicate notes piled up.

Source