it holds at scale
2.5×
Give both systems five times as much to remember and the lead holds.
01 · the difference
The same two messages, the same question, eight months later. One of these answers goes out to a stranger with your name on it.
what you told it
“I'm a software engineer at Stripe.”
“I left Stripe. I'm at Anthropic now.”
Existing memory stores
“A software engineer at Stripe.”
Eight months out of date. Both facts sit in the store with equal standing, so the agent returns whichever one the search happens to surface first, and it has no way to know the other one exists.
reMem
“A software engineer at Anthropic.”
Current, and it can show its work. Stripe was marked superseded the moment you said you left, and kept. Ask why and reMem points at the exact message that changed its mind.
02 · results
Head to head against existing systems on a benchmark built from facts that get corrected partway through.
3×
improvement over existing systems
45.0% against 15.0%. Three times as many questions answered correctly when the answer turns on a fact that was later corrected.
it holds at scale
2.5×
Give both systems five times as much to remember and the lead holds.
cost per memory
0
Zero cost to save or recall a memory.
every answer has a receipt
100% traceability
Everything reMem believes is grounded in what you said. Ask why, and it shows you the exact message.
03 · why it wins
Memory layers tend to do one of two things with a contradiction, and both lose something you need.
The store now holds the new value and nothing else. Nothing can explain why the answer changed, and if the update was wrong there is no earlier state to fall back to. You cannot audit what you cannot see.
Both values sit in the store with equal standing, so retrieval returns whichever one happens to embed closest to the query. The agent confidently reports a fact that stopped being true months ago.
reMem does neither. The old value stays on the record, marked, with the observation that replaced it still attached.
04 · how it works
01
Appends to the ledger. Synchronous, no model call, nothing to get wrong.
02
The model proposes belief updates, and a deterministic reducer decides what gets applied. The model never directly or freely mutates the underlying state.
03
Mutable, scoped, decaying, provenance-linked. A contradiction marks the prior belief superseded.
04
A scoped, ranked, compact context pack. It abstains rather than guessing.
An SQLite file and local embedder. No vector database, no graph database, no services to run. Everything on your machine, local by design.
A bundled viewer shows every belief the kernel holds, the value it replaced, and the messages both came from, updating live in your browser as your agent works.
05 · use it
npm i remem-kernelEight methods: observe, consolidate, decay, recall, why, beliefs, forget, export.
claude mcp add remem -- \
npx -p remem-kernel remem-mcpYour agent gets seven tools: recall, remember, beliefs and why for what memory holds; search, observation and history for what happened.
npm i -g remem-kernel
claude plugin marketplace add \
"$(npm root -g)/remem-kernel"
claude plugin install remem@rememRestart Claude Code and it starts remembering on its own. When the session ends, it consolidates what it learned, and the next session opens already knowing it.
npx -p remem-kernel remem-viewerPrints a live endpoint http://localhost:37800 so
you can see everything your agent remembers and believes.