rse_mem gives your LLM agent an exact, verifiable memory. Store a fact, get it back byte-for-byte — or an explicit MISS. No confabulation on what's stored. Ever.
When an LLM reconstructs a known fact from its weights, it can produce a confident, plausible, wrong answer. The fix isn't a bigger model — it's reading instead of reconstructing.
The model regenerates the fact lossily every time. When it's uncertain, it fills the gap with something that sounds right.
The fact is stored losslessly and returned exactly, with a proof — or an explicit MISS that tells the model to stop and ask.
Three properties make rse_mem different from embeddings, RAG, and vector search.
Facts are stored losslessly and reconstructed byte-for-byte. No embeddings, no summarization, no similarity guesswork — the value you get is the value you stored.
Every recall returns a cryptographic receipt. An independent verifier confirms the value against a published root — you don't have to trust the server, you verify it.
Ask for something that isn't stored and you get an explicit MISS — never an invented answer. "I don't know" becomes a first-class, reliable response.
rse_mem runs as an MCP server. Your agent calls it like any other tool.
Store a durable fact under a stable key — project.deadline, user.timezone, a decision, a result. It's kept exactly.
Before answering from memory, the agent calls memory_recall. A HIT returns the exact value plus a proof; a MISS says so plainly.
Any recall can be checked against the published root with memory_verify. A substituted or altered value cannot pass.
Register rse_mem as an MCP server in your agent client.
Runs locally. Your memory store stays on your machine.