A durable archive and a model context window answer different needs. Users may want a long account history for reading and continuity, while the current turn needs only a small, relevant portion.
Unlimited storage also increases deletion, export, privacy, and deduplication responsibilities. The system therefore keeps display history separate from the bounded context assembled for generation.
What was actually going wrong
The word “history” encouraged one array to serve as archive, memory database, and live prompt.
What I tried
- Appending the entire archive to every request
- Deleting older turns to protect prompt size
- Summarizing everything into one permanent paragraph
Cloud history remains browseable while recent slicing and scored retrieval choose what enters a prompt.
Why it worked
Storage can grow according to account needs without forcing inference cost and exposure to grow with it.
Remembering data and presenting data to a model are separate permissions.
Where EACI uses this today
Main accounts can retain conversation history while companion prompts stay bounded.
This journal covers real engineering on EACI Companion / The Veil. Companions include Caelum, Chad, Natalia, Atreus, Luna, Roxy, and Cael. Journal articles stay family-safe in content. See Privacy and Ethics.