Not a database, and not an LLM. A mind of linked scenes that learns by writing, in which the language model holds no authority.
What was known, and when
What it is
MEE — a Mind of Linked Scenes — learns without touching a single weight: it writes scenes, evidence, relations, rules, corrections and explicit decisions. It remembers what happened and how it knows it across millions of scenes, and it separates four things almost no system tells apart: what happened, what was believed then, what is known now, and what remains unknown. It can use a frozen language model as a linguistic organ, but that model is not memory, is not truth and does not decide: switching it off removes none of its essential cognitive capabilities.
Who it's for
Institutions that need a system able to answer for what it knows, not merely to sound convincing.
How it works
Scenes, not text
Nothing is stored as a block of text. Every event is broken down into typed entities, roles, changes, claims and evidence; the scene is a view over them.
Two clocks at once
Every claim keeps both when it was true and when it became known. Queries take an “as_of”: you can ask what was known at a specific past moment and what is known today, and those are two different answers, both correct.
Contradiction is not erased
When two sources contradict each other, MEE does not pick a winner: it marks the conflict and keeps it local. Correcting means appending a new record, never rewriting the old one. And when a premise is refuted, whatever was derived from it is withdrawn — without erasing the record that it was once believed.
Every conclusion can be rebuilt
“explain_claim” returns which rules were applied, in what order, and over what evidence. A conclusion with no rule, no inputs and no trace is not accepted as a factual answer: it is not flagged as doubtful, it is simply invalid.
Two implementations watching each other
A Python oracle fixes the meaning and a Rust core executes it. Both run the same scenarios and are compared field by field: any divergence is a bug, never “a version difference”.
Language is an organ, not the brain
The model never sees the ledger, touches the network, holds credentials or executes anything. It proposes; the core decides. Switched off, MEE keeps learning, reasoning and explaining through structured input.
Sovereignty and honest forgetting
It exports and imports onto another machine keeping identity and integrity. And it forgets for real: selective deletion and key destruction, without falsifying the history that remains.
Scope
The specification puts itself to the test: it sets ten conditions that, if they ever held, would invalidate the design. Four of them:
- That switching off the language model prevents querying, learning, reasoning, explaining or recovering.
- That a correction requires retraining or changing weights.
- That the state cannot be rebuilt from authenticated material.
- That the system only looks intelligent because it delegates memory or reasoning to the LLM.
Frequently asked
So why is there a language model inside at all?
For build cost, not for lack of capability in the core. Hand-coding every grammatical case so MEE could speak fluently would have eaten the time the core needs — and the core is where the value is. The model comes in as an engineering shortcut on the linguistic side, with its authority cut to zero. And the proof that it is a shortcut rather than a crutch is that model-off mode is verified by actually killing the process, not mocking it: with the model dead, MEE must still pass memory, rules, analogy, structured learning, explanation, backup and restore.
Frequently asked
Don’t Datomic, a temporal graph, or Postgres with history plus RAG already solve this?
It is a good question: for traceability of facts over time, a bitemporal store covers much of the ground. MEE goes one step further: here the unit is not the fact but the chain of justification — why something was concluded, on what evidence, believed at what moment, and what is withdrawn when a premise is refuted — and the whole system keeps working with the model switched off. For anyone who has to answer for why something was concluded, that chain is exactly what they need.
SPEC-MEE-1.0 authorises implementation and testing; declaring production requires certifying its 361 requirements one by one, and that certification is under way. 30 of the 46 operations already run on the Rust core.