03 AI architecture · Evaluation
Semantic memory
Memory that cites its source.
The question
What if a model compiled what it knows once — and every answer stayed tied to its evidence?
- Role
- Concept originator; research direction
- Years
- 2026
- With
- H3M research
- Status
- Research
- Disciplines
- AI architecture, Semantic memory, Evaluation
- Constraint
- Language models rebuild meaning from raw text on every query, and their answers drift from their sources.
Method
01/04
Source material becomes typed semantic objects, relations, and byte-exact source spans — stored, versioned, and reusable.
Invented text, to show the step
02/04
A query wakes only the small subgraph it needs.
Invented text, to show the step
03/04
Deterministic operators do the exact work; a recurrent reasoner handles what binding can’t resolve.
Invented text, to show the step
04/04
Every durable claim traces back to exact evidence, and ambiguity is preserved when context is insufficient.
Invented text, to show the step
Results
- 97%CSFN Phase 5 report: held-out reasoner benchmarkOne synthetic microworld. A zero-parameter baseline scores 31/37, so the learned margin is 31→36. Not evidence of general reasoning.36 of 37 held-out reasoning tasks, matching a frozen LLM baseline.One synthetic microworld. A zero-parameter baseline scores 31/37, so the learned margin is 31→36. Not evidence of general reasoning.
- 5,152CSFN Phase 5 report: held-out reasoner benchmarkOne synthetic microworld. A zero-parameter baseline scores 31/37, so the learned margin is 31→36. Not evidence of general reasoning.Active parameters in the recurrent reasoner; a further 32 are never read.One synthetic microworld. A zero-parameter baseline scores 31/37, so the learned margin is 31→36. Not evidence of general reasoning.
- 31 → 36CSFN Phase 5 report: held-out reasoner benchmarkOne synthetic microworld. A zero-parameter baseline scores 31/37, so the learned margin is 31→36. Not evidence of general reasoning.Held-out tasks at one reasoning step, then at two or more — the same frozen weights, iterating.One synthetic microworld. A zero-parameter baseline scores 31/37, so the learned margin is 31→36. Not evidence of general reasoning.
Latest · Oct 2026
From a controlled test world to real text.
CSFN is in active development. The first result was measured in a controlled test world; the work now takes the same discipline to real text, where the hard part is reading a sentence exactly. Each round is frozen in advance with its own pass mark, and a candidate is adopted only if it clears it.CSFN development record, 15–16 October 2026Work in development. The count is the whole automated suite at commit a523320b29 (5,142 passed, 2 expected failures, fixed seed). The real-text rounds are exploratory: no result from them is claimed here, and the earlier test-world result is unchanged.
- DoneA controlled test worldThe approach, measured against a baseline on questions it had not seen.
- NowReal textFirst rounds on sentences from an open academic corpus: each passage tied to its exact place in the original, and what a sentence leaves unsaid kept as unknown.
- NextThe whole chainSource to answer, checked end to end: nothing answers unless every step back to its words holds.
Carried forward
Memory is only useful when it can show its work.
Credits
- Ryan BrantConcept originator; research direction
- H3MStudio