structuremappingmemory 1.0.0

SMA-1: Structure-Mapping Agentic Memory — a structure-mapping retrieval memory that grounds language models in curated ontologies

SMA-1: Structure-Mapping Agentic Memory

Structure-mapping memory grounds language models in curated ontologies. SMA is a retrieval memory that grounds a generalist LLM in a curated, expert-maintained ontology, retrieving by logical structure — subsumption (is-a) and higher-order relations — that vector RAG and knowledge graphs discard. Holding the language model and prompt fixed and swapping only the memory, the SMA-grounded agent becomes more accurate, more selective, and more structurally attributable than vector RAG: it cites its evidence structurally, abstains when nothing grounds the case, and flags novelty — capabilities flat retrievers structurally lack. On the real-domain benchmarks the measured advantage is carried by bounded is-a subsumption plus information-content weighting (the higher-order relational machinery is decisive only on a synthetic structure-only control), and it vanishes where no discriminative ontology applies.

A structure-mapping memory over a curated ontology grounds a generalist LLM by the subsumption hierarchy and rarity weighting that flat retrievers wash out — and the structural alignment itself is what survives when there is no surface signal at all.

---

What it does

Capability	SMA	Vector RAG	Knowledge graph
Rare / long-tail retrieval (rarity-weighted)	✅	✗ averages rarity away	✗
Cross-vocabulary / zero-lexical-overlap match	✅	✗	partial
Subsumption (is-a) reasoning	✅	✗	✗ no ascension
Structural citation (checkable provenance)	✅	✗	partial
Calibrated abstention	✅	✗ cosine always "high"	✗
Novelty / expectation-violation flag	✅	✗	partial

One universal loader ingests any open OBO/OWL/STIX/CPC ontology (is-a hierarchy plus typed relations, no OWL reasoning); a registry + domain router select across domains. We mount eleven ontologies spanning seven knowledge areas (≈594,000 concepts) behind one shared retrieval core plus thin per-domain data adapters.

Headline results (honest)

On a memory-swap benchmark where only the retriever varies, SMA beats a strong RAG + KG baseline suite on the rare/long-tail slice across five domains:

Domain	SMA tail top-5	best RAG	Δ	Holm-significant
Medicine (HPO/MONDO)	0.949	0.606	+0.333	yes
Finance (US-GAAP)	0.418	0.231	+0.167	yes
Genomics (GO)	0.849	0.682	+0.156	yes
Legal (CPC, all-query)	0.941	0.870	+0.064	yes
Cyber (ATT&CK)	0.766	0.749	+0.073	directional*

* Cyber survives Holm across domains but not a conservative Bonferroni-over-baselines selection correction (p=0.035 → 0.17); reported as directional. End to end (medicine, LLM-in-the-loop), the SMA-grounded agent is the most accurate (0.342 vs dense 0.100 vs closed-book 0.017), most faithful in citation, and uniquely separates known from unknown entities (grounding AUROC 0.79 vs vector RAG's near-chance 0.55).

Where it does not help (reported, not hidden): SMA ties a bespoke ontology information-content oracle (Phenomizer) rather than beating it; and on flat single-record tabular prediction (hospital readmission, credit-card fraud, the Elliptic Bitcoin graph-fraud probe) there is no discriminative subsumption ontology to mount, so SMA reaches only parity and a flat logistic-regression baseline wins. The advantage is specific to structure.

How it works

1. Adapter — a universal loader parses any open ontology into a normalized graph, mounts the is-a hierarchy as an ascension lattice (specific terms match general ones with penalty ρ^dist) and the typed relations as higher-order cases, and registers them behind a per-query domain router.
2. Matcher — a case (functors over a subject) is screened by a certified MAC admissible-content bound, aligned by best-first structure-mapping (FAC/SME), and scored by a rarity-weighted information-content (−log₂ p) surprisal scorer.
3. Decision — the alignment yields a structural citation, a calibrated abstention, or a novelty flag (SAGE expectation violation).

Quick start

python -m venv .venv && . .venv/bin/activate
pip install -e ".[encoders,eval]"
pytest -m "gate_G0 or gate_G1 or gate_G2 or gate_G3 or gate_G4 or gate_G5 or gate_G6"
sma ui                       # FastAPI + Gradio demo

Raw datasets are not tracked. Fetch checksum-verified external data with:

python scripts/fetch_datasets.py --manifest data/manifests/datasets.json

Reproduce the confirmatory tables/figures from committed CSVs:

python scripts/figures_paper.py      # main-text figures
python scripts/figures_ed.py         # Extended Data figures

All result numbers in the paper trace to reports/confirmatory/*.csv; runs are deterministic (PYTHONHASHSEED=0, paired bootstrap seeded at 12345).

Repository layout

Path	Contents
sma/ontology/	universal ontology adapter (loaders, mount, registry, router) — frozen adapter-v1
sma/eval/agentic/	memory-swap retrieval harness + baseline arms
sma/eval/agentic_qa/	end-to-end LLM-QA agent (pre-registration v2)
scripts/	dataset fetch, figure/table generators, ablations
reports/confirmatory/	all result CSVs (every paper number sources here)
paper/manuscript/	sma_nature_mi.tex — the manuscript (Nature sn-jnl template)
release/	model card, dataset card, HuggingFace Space

Paper & citation

The manuscript (paper/manuscript/sma_nature_mi.tex, intended for Nature Machine Intelligence) is built on the official Springer Nature sn-jnl template. To cite this software, see CITATION.cff.

License

Apache-2.0. All mounted ontologies are openly licensed and version-pinned; we redistribute only derived gold labels, never restricted source records (see Extended Data and release/model_card.md).