What sepiace does
sepiace gives your agent long-term memory through an MCP tool that takes one natural-language sentence. This page covers its features, how a request is handled, and its KiokuBench results.
Features
The server decides what to do
A judgment network on the server decides whether to store, recall, search, or weaken memories, and combines them when needed. Your agent passes a sentence and spends no tokens building structured arguments.
Memories are your own words
The text of each memory is cut from the sentence in the request. No LLM generates or rewrites it. Responses contain only memory text, with no IDs, timestamps, or scores.
Scores follow a forgetting curve
Each memory's score is computed at read time with the FSRS forgetting curve. Memories that go unused fade gradually, and each recall makes them stick.
Memories are separated per user
You sign in with a six-digit code or a link sent to your email. Every database query and every vector search is scoped to the signed-in user.
Memories stay in your language
Because the text is cut from your request, a memory stays in the language you wrote it in. In KiokuBench v0.1, none of sepiace's memories were written in a different language from the conversation.
How it works
Your agent sends one sentence
For example: “The user prefers short answers, but wants detailed code reviews.” The MCP tool takes natural language only.
The judgment network decides
The network on the server chooses the operations. When it is unsure, it asks again with a deeper layer instead of guessing.
The text is cut out and stored
The memory text is cut from the sentence and stored under the signed-in user. Scores are computed from the forgetting curve when memories are read.
Only memory text comes back
Your agent receives the memory text and nothing else.
KiokuBench results
KiokuBench checks what a memory system remembered, replaced, and forgot by comparing the stored memories against a ledger of true facts. Below are the totals for the route where the raw conversation is passed in as is.
| Metric | sepiace, raw conversation | Supermemory, raw conversation |
|---|---|---|
| Recall | 99% | 98% |
| Update applied | 99% | 100% |
| Old version kept | 1% | 5% |
| Forget violation | 0% | 78% |
| Noise | 13% | 20% |
| Language mismatch | 0% | 42% |
| Check questions | 97% | 88% |
| Compression | 1.29 | 2.30 |
sepiace matched or beat Supermemory on every metric except one: when a change is signaled explicitly, sepiace kept the old version of a changed fact more often, at 22% against 11%. This is a single run on 10 personas.
Connect sepiace to your agent
Register one URL in your MCP client and sign in with your email.