How sepiace remembers and forgets
This page explains how sepiace decides what to remember, recall, and forget, and how it treats your memories over time. It describes the design, not every detail of the implementation.
A layered judgment network
sepiace does not leave your request to one large LLM prompt. It stacks many small judgments in layers, and together they decide whether to remember, recall, or forget.
When a layer is unsure, a deeper layer asks the question again, narrowing the scope or adding background knowledge.
If the answer is still unclear, sepiace chooses the side that does not damage your memories. For a forget request, once it is clear what should be forgotten, it chooses the side that does not keep it.
Not an LLM summary
The text of each memory is cut only from the sentences you pass in. No LLM paraphrases, summarizes, or adds to it.
What comes back is memory text and nothing else.
Dreaming
Once a day, sepiace spends time dreaming, which is when it tidies up memories. Sentences that ended up scattered are gathered back into the memories where they belong.
Dreaming does not rewrite sentences either, and each sentence keeps how well it has settled in memory.
A forgetting curve
Memories that go unused become gradually harder to recall, and each recall makes them stick.
Scores are computed from the forgetting curve at the moment memories are read.
Facts that change, and things to forget
When a fact changes, sepiace replaces the old version.
When you ask it to forget something, sepiace removes what you named, and from then on it never appears in recall or search. The removed content is kept as a logically deleted record and is physically deleted when you request full deletion. If it is not clear what should be forgotten, it forgets nothing.
In your words, and yours alone
Because the text is cut from what you wrote, memories stay in the language of the conversation.
Memories are kept separate for each user. Every lookup and every search covers only the signed-in user’s memories.