Research · AI · Machine Learning · Neural Media · Research
Semantic Memory: AI That Cites Its Source
Language models rebuild meaning from raw text on every query, and their answers drift from their sources. CSFN, an H3M research project, compiles knowledge once into typed objects tied to exact evidence. In a controlled test world, a 5,152-parameter reasoner matched a frozen LLM baseline on 36 of 37 held-out tasks.
· Ryan Brant
Ask a large language model the same question twice and it rebuilds its understanding from raw text both times. That is expensive, and it means answers can drift away from the sources they came from. For AI that people rely on, in contracts, research, or the worlds we build in Portals, an answer is only as good as the evidence you can point to.
CSFN is H3M's research into a different kind of memory. The question: what if a model compiled what it knows once, and every answer stayed tied to its evidence?
Compile once
Source material is read once and turned into typed semantic objects (people, places, dates, amounts), the relations between them, and byte-exact spans pointing back to where each one appears in the original text. The result is stored, versioned, and reusable.
Activate little
A question doesn't need everything the system knows. A query wakes only the small subgraph it needs, and everything else stays asleep.
Reason small
Deterministic operators do the exact work, such as looking up, subtracting, and comparing, and they give the same answer every time. A small recurrent reasoner handles what binding alone can't resolve, such as working out which item a pronoun refers to, by looping through a few passes.
Verify everything
Every durable claim traces back to exact evidence. When the context is insufficient, ambiguity is preserved: the open question is kept, with its candidates listed, rather than guessed.
The first result
In a controlled test world, the approach was measured against a frozen large language model on reasoning tasks it had not seen:
- 97%: 36 of 37 held-out reasoning tasks, matching the frozen LLM baseline. - 5,152 active parameters in the recurrent reasoner. - 31 to 36: held-out tasks solved at one reasoning step, then at two or more steps, with the same frozen weights iterating.
The scope matters, so here it is in full: this is one synthetic microworld. A zero-parameter baseline scores 31 of 37, so the learned margin is 31 to 36. It is not evidence of general reasoning. What it shows is that a very small, inspectable reasoner can close the gap on the cases that need it, while everything it says stays tied to its source.
In development: from a test world to real text
CSFN is in active development. The work is now taking the same discipline to real text, where the hard part is reading a sentence exactly. The first rounds use sentences from an open academic corpus, with each passage tied to its exact place in the original and anything a sentence leaves unsaid kept as unknown.
Each round is frozen in advance with its own pass mark, and a candidate is adopted only if it clears it. As of 15 October 2026, 5,142 automated checks pass across the project. Next is the whole chain, from source to answer, checked end to end: nothing answers unless every step back to its words holds.
Memory is only useful when it can show its work.