I Built an MCP Memory Server Where Code, Not the Model, Decides What Gets Stored
There's a hole in almost every AI memory library, and once you see it you can't unsee it: the model that hallucinates is also the thing you've put in charge of the record. You hand an LLM a conversation and ask it to remember what matters. It reads, it decides, it writes its interpretation into long-term memory as if it were a fact. Three sessions later your app confidently "knows" something the…
There is a flaw in most AI memory libraries. Even after an LLM reads a conversation and tries to remember the important parts, it can still invent facts that aren't there. The author spent a year trying to solve this problem. The solution they built, which they just published, inverts the usual setup. The model no longer decides what gets stored; instead, deterministic code decides.
This model proposes claims, and then deterministic code verifies if the evidence supports those claims before committing them to long-term memory. If the evidence doesn't support the claim, it is refused and not stored. The author also provides a way to audit the provenance of stored facts by storing the exact byte range of the source document where the claim was found.
If the source document changes later, the stored receipt will fail verification, alerting the system to the change. In cases where a claim is outside the scope of stored facts, the system abstinence instead of fabricating an answer, telling the user exactly what it doesn't know. The system also offers a way to delete facts, with a signed certificate proving the deletion, ensuring that deleting one fact won't accidentally affect other related facts.
While the system is solid, the code is still young and the author admits to finding bugs within minutes of installing it.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
