Mathematical Foundation

Geometry,
not guesswork.

Three public arXiv preprints describe versioned research experiments. In the shipped path, dense candidates use cosine similarity; Fisher-derived terms can inform later scoring. Product behavior is governed by release code, tests, and measured artifacts rather than a component theorem.

The boundary

Where information geometry fits.

Cosine similarity is the shipped dense candidate metric. It measures embedding direction and does not, by itself, encode the confidence stored with an assertion. That limitation does not make cosine invalid; it defines what later scoring stages may add.

Score meanings stay separate. Query-relative relevance, internal ranking utility, stored-memory confidence, trust, and answer confidence are different values. Score Contract v2 preserves those distinctions.

Fisher-derived research terms can influence post-candidate scoring when their required state is available. They do not replace every cosine path, and they do not turn rank into answer probability. Until a held-out calibrator is frozen, answer_confidence remains null.

Field Meaning V4 status
relevance_score query-relative relevance bounded 0..1
ranking_score internal ordering utility not a probability
memory_confidence stored assertion field separate from relevance
answer_confidence calibrated answer probability null / uncalibrated
Canonical response: score_contract_version: "2"
Calibration status is explicit. V4 does not present retrieval ordering as answer probability.
01 — Retrieval Metric
Fisher-Rao Distance · arXiv:2603.14588
dFR(p, q) = arccos( Σ √(pi · qi) )

Where p and q are probability distributions over memory confidence scores. The sum Σ is taken over all memory dimensions i. The arccos maps the result to angular distance on the unit sphere — the natural geometry of probability distributions under the Fisher information metric.

This is the geodesic distance on the statistical manifold — the shortest path between two probability distributions, measured along the curved surface they inhabit. Not the Euclidean shortcut through the void. The manifold curvature is determined by the Fisher information matrix, which encodes how much information each dimension carries. High-confidence dimensions curve the space more steeply. Low-confidence dimensions barely curve it at all.
  • Symmetric dFR(p, q) = dFR(q, p)
  • Triangle inequality dFR(p, r) ≤ dFR(p, q) + dFR(q, r)
  • Degenerate at identical distributions dFR(p, p) = 0
  • Bounded 0 ≤ dFR(p, q) ≤ π / 2
2603.14588 — V3 SLM: Information-Geometric Agent Memory →
02 — Lifecycle Model
Riemannian lifecycle research · arXiv:2604.04514
γtγ = 0     ·     Expp(v) = γ(1)

The preprint studies geometry-inspired forgetting, consolidation, and cognitive quantization. The equations describe the research model evaluated there; they are not a complete specification of every current runtime transition.

The released lifecycle also contains explicit policy, usage, recency, thresholds, and maintenance code. Those inspectable mechanisms—not this equation alone—define shipped behavior.

A mathematical property of one component does not guarantee end-to-end retention or recall. V4 lifecycle claims require update, supersede, archive, restore, forget, deletion, and stale-representation tests across the frozen artifact.
2604.04514 — The Living Brain →
03 — Compression Theory
Compression evidence boundary
H(X|Y) H(X)     ·     I(X;Y) 0

These are Shannon's fundamental information inequalities. H(X|Y) is the conditional entropy of prompt X given prior context Y — the irreducible information content that cannot be compressed away without information loss. I(X;Y) is mutual information — the quantity of information that X and Y share, which is always non-negative.

These inequalities describe theoretical limits; they do not prove a compression ratio, fidelity result, or provider saving for the shipped implementation. Product claims require a declared corpus, mode, tokenizer, round-trip test, latency measurement, and raw artifact.

Safe mode preserves JSON and code and may produce no reduction. Aggressive prose compression is separately opt-in and lossy. Shannon inequalities cannot be used to turn either behavior into a generic product-wide percentage.
  • Structured payloads (JSON, code) Measure reduction and round trip per corpus
  • Unstructured prose Aggressive path is opt-in and lossy
  • Byte-exact reversible on structured paths Verify byte invariants in release tests
compression — live example
# Check prompt compression on a structured payload $ slm compress --status --payload schema.json ↳ record input and output bytes/tokens for the supplied corpus safe path: verify structured-data invariants and byte round trip provider cache: measure separately from compression
See the shipped compression boundaries →
The field

Why information geometry?

Background · Amari & Rao, 1945–1985

A geometry built for probability.

Information geometry, developed by Shun-ichi Amari and C. R. Rao, is the study of probability distributions as geometric objects. The key insight: probability distributions do not live on a flat plane. They live on a curved manifold — the statistical manifold — where the natural notion of distance is the Fisher information metric, not Euclidean distance. Amari's 1985 monograph unified differential geometry and statistics into a single framework that lets you reason about distributions the same way classical geometry reasons about shapes.

Why it fits agent memory

Memories are distributions, not points.

The research asks whether probability-aware geometry can add useful signals beyond embedding similarity. In the current implementation, dense candidates still use cosine similarity and Fisher-derived terms may modify later scoring. Whether that improves retrieval is an empirical question for a frozen corpus and trace, not an automatic consequence of the theory.

Public Research

Three arXiv preprints.

The research claims on this page are traceable to public arXiv preprints. Read the methods, reproduce the results, and cite the work.

01 · Retrieval metric
dFR(p, q) = arccos( Σ √(pi · qi) )

Fisher-Rao Distance

Historical information-geometric retrieval experiments. Current dense candidate generation uses cosine; Fisher-derived terms may inform later scoring.

arXiv:2603.14588 →
02 · Trust and privacy
γtγ = 0  ·  Expp(v)

Bayesian Trust Defense

Historical evaluation of trust, privacy, search, graph, and multi-agent memory-poisoning defenses.

arXiv:2603.02240 →
03 · Forgetting and quantization
H(X|Y) ≤ H(X)  ·  I(X;Y) ≥ 0

The Living Brain

Historical forgetting, cognitive quantization, and multi-channel retrieval experiments. Not a generic prompt-compression proof.

arXiv:2604.04514 →