MCP server · Claude Code, Codex, any MCP client
Your agent forgets you every session. sepiace doesn't.
Long-term memory for Claude Code, Codex, and any MCP client. It keeps what you tell it in your own words, and brings it back when it matters.
Get started freeSee the benchmark
- Claude Code
- 18.9 → 60.6
- Codex
- 11.4 → 66.1
Session · Oct 3
youMia's coming over on the 20th.
youRaw apples are off-limits for her, but cooked is fine.
agentNoted.
New session · Oct 29context: empty
youPlan a dessert for Mia's birthday — mind her allergy.
sepiace.memory("What should I know about Mia's allergy?")
→ “Raw apples are off-limits for her, but cooked is fine.”
rewritten memory: “Mia is allergic to apples.”
a rewrite drops the ‘but’
agentA baked apple tart — cooked apples are fine for Mia.from memory
Kept the clause after ‘but’: Claude Code + sepiace 80% · built-in memory 3% — KiokuBench v1 r1
Against Claude Code and Codex built-in memory
We gave Claude Code and Codex the same conversations for 12 people, once with only their built-in memory and once with sepiace connected. Then, in a new session, we asked about each person.
- *1 Codex creates memories only from interactive sessions, so for its built-in memory we drove Codex through its interactive CLI screen.
Connecting sepiace raised the Index with both agents. Neither is close to a perfect score, and Replace and Combine are still low. With Claude Code, taking in the conversations used a little more usage and time with sepiace.
It doesn't just store it. It gets it right.
Share of correct answers in four KiokuBench v1 r1 scenes. sepiace raised the first three a lot; the fourth is still ahead of us.
Against other memory layers
KiokuBench checks what a memory system remembered, replaced, and forgot against a ledger of true facts, and combines four pillars into an Index from 0 to 100, where higher is better. Below is the route where the raw conversation is passed in as is.
- sepiace96.1sepiaceIndex: 96.1Remember: 98.7Replace: 98.0Forget: 100.0Tidy: 88.3
- Supermemory58.9SupermemoryIndex: 58.9Remember: 95.3Replace: 95.0Forget: 22.0Tidy: 60.5
- ByteRover39.2ByteRoverIndex: 39.2Remember: 72.3Replace: 19.0Forget: 62.0Tidy: 27.7
- Mem08.5Mem0Index: 8.5Remember: 96.3Replace: 0.0Forget: 0.0Tidy: 55.0
- Store everything8.0Store everythingIndex: 8.0Remember: 99.0Replace: 0.0Forget: 0.0Tidy: 41.7
Hover over or focus a bar to see its four pillars. The store-everything baseline is shown in gray for reference.
sepiace scored highest on the KiokuBench Index and on all four pillars. Supermemory, Mem0, and ByteRover left much of what they were asked to forget, which pulled down their Forget pillar. This is a single run on 10 personas.
These KiokuBench v0.1 scores cannot be compared with the v1 scores above.
KiokuBench v1 r1 tables, how we measure, conditions, and limitations
The detailed KiokuBench v0.1 tables for all eight metrics, conditions, and limitations
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, so your agent spends no tokens building structured arguments.
A judgment network decides
A network on the server chooses whether to remember, recall, search, or forget, and combines them when needed. When it is unsure, it judges again at a deeper layer instead of guessing.
Your words are kept as they are
The memory text is cut from the request. No LLM writes or rewrites it, so it stays in the language you wrote it in. In KiokuBench v0.1, none of sepiace's memories were stored in a different language, while Supermemory stored about 80% of Japanese input as English. Memories are kept per user: you sign in with a six-digit code or a link sent to your email, and every query and vector search is scoped to the signed-in user.
Only the memory comes back
Recall returns the memory text and nothing else: no IDs, timestamps, or scores. Scores follow the FSRS forgetting curve, so memories that go unused fade, and recalled ones settle in.
Connect with one command
Run the command below in your terminal. It finds Claude Code, Codex, Cursor, VS Code, and OpenCode on your machine, connects the ones you choose, and signs you in with your email.
npm create sepiaceStart free
From $9 a month for 1,000 credits (about 200 saves). Your first month is free, no card.