sepiace

Long-term memory for AI agents. Your agent passes one natural-language sentence, and sepiace remembers, recalls, or forgets.

Quick start

npm create sepiace

It finds Claude Code, Codex, Cursor, VS Code, and OpenCode on your machine, connects the ones you choose, and walks you through signing in. Then start a new session and ask your agent:

• "Remember that I prefer pnpm over npm."

• "What did we decide about the release schedule?"

• "Forget my old home address."

To bring over what your agent already remembers, run /sepiace:migrate in Claude Code, or ask other clients to use the sepiace-migrate skill. It shows you what it found and sends nothing until you agree.

Claude Code:

/plugin marketplace add osaxyz/sepiace
/plugin install sepiace@sepiace

Codex:

codex plugin marketplace add osaxyz/sepiace
codex plugin add sepiace@sepiace
codex mcp login sepiace

Other MCP clients: add https://hi.sepiace.io/api/v1/mcp as a remote MCP server (Streamable HTTP). The client signs in through OAuth in the browser.

Related MCP server: friday

Technology

Your agent writes what it wants in plain words. A judgment network on the server decides whether to save, recall, search, or forget, and combines them when needed. The response contains only memory text, with no IDs, dates, or scores, so the agent spends no tokens building structured arguments.

Each memory is cut from the sentences in the request. No LLM writes or rewrites it, and when something becomes outdated, only that sentence is replaced. A memory stays in the language you wrote it in.

Each memory's score follows the FSRS forgetting curve and is computed when it is read. Memories that go unused fade, and once a memory's recall probability drops below 50%, it no longer appears in search. Each recall makes it stick.

Memories are kept separately for each account, and every database query and vector search is scoped to the signed-in user. The console at hi.sepiace.io lists every request with the memories it changed, and a client you disconnect there stops working right away.

KiokuBench checks what a memory system actually stored: whether it kept the facts, replaced outdated ones, forgot what it was asked to forget, and stayed tidy. The KiokuBench Index combines these four pillars into one number from 0 to 100.

KiokuBench v0.1, 10 personas in English and Japanese, raw conversation passed in as is. sepiace was measured on October 14, 2026, and the others between September 27 and 30, 2026. KiokuBench was designed and run by the team that builds sepiace. The method, the eight metrics, the conditions for each system, and the limitations are on sepiace.io.

Specification

It never changes the local files.

The CLI lives in cli/ and is published to npm as create-sepiace.