Skip to content

Research Foundations

Aevum is built on established science, open standards, and published research. This page documents the sources that informed each major architectural decision.

The synthesis — the five functions, the barrier hierarchy, the complication model, the governed membrane, and the sigchain design — is original to Aevum. None of these sources describe Aevum or anything equivalent to it. What they provide is the scientific and engineering foundation that each individual decision stands on.

We credit these works because intellectual honesty demands it, and because users of a governance system deserve to know that its design choices are grounded, not arbitrary.


Cognitive Science and Memory

Power-Law Forgetting — Edge Weight Decay

The knowledge graph uses power-law decay for edge weights rather than exponential decay. This decision is grounded in two converging bodies of research.

Jost's Law (Jost, 1897) observed that older memories are forgotten more slowly than newer ones of equal current strength — a phenomenon exponential decay cannot model but power-law decay naturally captures.

Wixted, J.T. (2004). The psychology and neuroscience of forgetting. Annual Review of Psychology, 55, 235–269. This review established that retention across nearly every studied domain follows a power function, not an exponential, and that the power-law is one of the most reliable findings in memory science. Aevum's decay function weight × (1 + days)^(-d) implements this directly. Sensitive edges use a steeper exponent (× 1.4) following the principle that emotionally significant material exhibits different decay dynamics.

BFS Depth Limit — Why Depth 2

Graph traversal in Aevum stops at depth 2 by default. Going deeper produces noise, not signal.

Balota, D.A. & Lorch, R.F. (1986). Depth of automatic spreading activation: Mediated priming effects in pronunciation but not in lexical decision. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12(3), 336–345. This study confirmed that automatic spreading activation in human semantic memory effectively operates at one to two steps. Depth 3 associations produce negligible facilitation. The BFS depth limit in Aevum is not an engineering convenience — it mirrors the empirically established reach of human associative memory.

Fan Effect — Activation Dilution at Branching Nodes

Highly connected nodes receive an activation penalty proportional to their degree. This is the fan effect from ACT-R.

Anderson, J.R. (1983). The Architecture of Cognition. Harvard University Press. ACT-R (Adaptive Control of Thought — Rational) established that retrieval time and activation strength degrade as the number of associations from a concept increases. A concept connected to many others is harder to retrieve precisely because activation spreads across more paths. Aevum applies this as an explicit penalty at high-degree nodes during traversal.

Complementary Learning Systems — Dual Memory Model

The separation between fast episodic storage (the sigchain) and slow semantic storage (the knowledge graph) reflects the Complementary Learning Systems (CLS) theory of human memory organisation.

McClelland, J.L., McNaughton, B.L., & O'Reilly, R.C. (1995). Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory. Psychological Review, 102(3), 419–457. CLS theory describes a fast-learning hippocampal system for specific episodes and a slow-learning neocortical system for statistical regularities. Aevum separates these concerns deliberately: the episodic ledger records specific events with full fidelity; the knowledge graph accumulates weighted semantic relationships over time.


Graph Theory and Information Retrieval

Personalized PageRank — Context Traversal

Aevum uses Personalized PageRank (PPR) for graph traversal rather than simple BFS or cosine similarity.

Page, L., Brin, S., Motwani, R., & Winograd, T. (1999). The PageRank citation ranking: Bringing order to the web. Stanford InfoLab Technical Report. PPR seeds the random walk from the query concept rather than the full graph, producing a relevance ranking that naturally decays with graph distance while respecting the full topology. For sparse, typed knowledge graphs, PPR consistently outperforms flat similarity retrieval on precision at small k.

PPMI — Concept Importance Scoring

Church, K.W. & Hanks, P. (1990). Word association norms, mutual information, and lexicography. Computational Linguistics, 16(1), 22–29. PPMI identifies concepts that are meaningfully associated above what chance predicts rather than merely frequent. Aevum applies PPMI as an IDF-like importance weight during context assembly, suppressing high-frequency but low-information concepts.


Cryptography and Audit Integrity

Ed25519 — Signing

Every AuditEvent in the sigchain is signed with Ed25519.

Bernstein, D.J., Duif, N., Lange, T., Schwabe, P., & Yang, B.Y. (2011). High-speed high-security signatures. Journal of Cryptographic Engineering, 2(2), 77–89. RFC 8032Edwards-Curve Digital Signature Algorithm (EdDSA). IETF, 2017. Ed25519 was chosen over ECDSA P-256 for its resistance to implementation side-channel attacks, fixed-size 64-byte signatures, and deterministic signing (no per-signature randomness required).

SHA-3 (SHA3-256) — Hash Chaining

NIST FIPS 202SHA-3 Standard: Permutation-Based Hash and Extendable-Output Functions. NIST, 2015. SHA3-256 (Keccak) was standardised as a complement to SHA-2 with a fundamentally different internal construction (sponge function), providing security against structural attacks that could theoretically affect SHA-2 family members.

TOCTOU Vulnerabilities in Agent Systems

The Context Witness (Phase 12a) was designed in response to published research on a specific class of vulnerability.

Lermen, S. et al. (2025). Mind the Gap: Time-of-Check to Time-of-Use Vulnerabilities in LLM-Enabled Agents. arXiv:2508.17155. This paper formally defined and benchmarked TOCTOU vulnerabilities in LLM agent pipelines. Aevum's witness mechanism directly addresses the State Integrity Monitoring defense described in this work.

OpenPort Protocol Specification (2026). arXiv:2602.20196. The witness design follows the State Witness profile: a minimal non-secret snapshot of relevant resource state, bound to the request at preflight, revalidated at execution time with fail-closed semantics.

Optimistic Concurrency — ETag Pattern

RFC 9110HTTP Semantics. IETF, 2022. Rather than holding a database lock for the duration of human review, Aevum captures a fingerprint at read time and uses it as a precondition at commit time — the ETag/If-Match pattern from HTTP concurrent editing.


Agent Security Research

Memory Poisoning — Barrier 3 and Policy-Mediated Writes

OWASP Top 10 for Large Language Model Applications (2025). Open Web Application Security Project. OWASP Agentic AI Security Top 10, ASI06: Memory and Context Poisoning (2025). Open Web Application Security Project. ASI06 defines memory poisoning as the injection of malicious or misleading content into an agent's persistent memory stores. The Aevum governed membrane (Barrier 3: consent enforcement, Barrier 5: provenance chain) ensures no data enters the knowledge graph without a valid consent grant and full chain of custody.

Agent Autonomy Levels

Levels of Autonomy in AI Systems. DeepMind Safety Research. The L1–L5 taxonomy distinguishes operator-controlled (L1) from fully autonomous (L5) agent behavior. Aevum enforces the declared level through policy rules: an agent at L3 cannot self-approve actions that require L1 authorization.


Open Standards

Standard Body Role in Aevum
OWL 2 W3C (2012) Knowledge graph ontology
SPARQL 1.1 W3C (2013) Graph query language
SHACL W3C (2017) Constraint validation
R2RML W3C (2012) Database-to-RDF mapping
BFO 2.0 / ISO/IEC 21838-2:2021 ISO Top-level ontology
RFC 9457 IETF (2023) HTTP error format
RFC 8032 / Ed25519 IETF (2017) Event signing
NIST FIPS 202 / SHA-3 NIST (2015) Hash chaining
UUID v7 IETF draft Time-ordered identifiers
OpenID Connect 1.0 OpenID Foundation Identity federation
MCP (Model Context Protocol) Anthropic Agent tool interface

Why SHACL over ShEx

W3C Recommendation status, native SPARQL extensibility, and superior tooling support across the RDF ecosystem. ShEx is a community specification without equivalent tooling depth.

Why BFO 2.0

The only top-level ontology published as an ISO standard, adopted by over 650 projects, and mandated by the US Department of Defense (January 2024).


Regulatory Frameworks

Aevum does not claim compliance — it produces the evidence that compliance work requires.

Framework Relevance
EU AI Act (2024), Article 12 Sigchain provides tamper-evident technical documentation of every AI decision
ISO/IEC 42001:2023 Complication lifecycle and governance controls
GDPR, Article 17 Consent revocation is immediate; erasure is recorded
SOC 2 Type II, PI1.2 Provenance chain and classification ceiling
HIPAA Classification ceiling at level 3 prevents unauthorized access

Shapiro, M., Preguiça, N., Baquero, C., & Zawirski, M. (2011). Conflict-free Replicated Data Types. SSS 2011, LNCS 6976, 386–400. The OR-Set CRDT allows adds and removes to commute — concurrent revocations and grants always converge to the same state regardless of the order they are applied, making the consent ledger safe for multi-node federation without a central coordinator.


A Note on Synthesis

None of the works cited above describe anything like Aevum. Wixted describes forgetting curves; he does not describe a governed AI context kernel. The OpenPort Protocol describes a governance specification; it does not describe a Python implementation with a sigchain and complication framework.

The synthesis — combining these independently published findings and standards into a coherent architecture for governed AI context — is the original work. This page exists to make the inputs to that synthesis transparent, not to diminish it.