Research
Mnemoverse memory architecture is grounded in peer-reviewed research — the SLoD paper was accepted at the GRAAI workshop (1st Workshop on Graphs Across AI, IEEE WCCI 2026). This page collects the science behind the engine's design. Analysis, market landscape, and longer-form writing live in the Library.
Articles in this section
- AI Agent Memory Crisis: Why Bigger Context Fails: why enlarging the context window does not give an agent memory.
- AI Agent Memory: The 2026 Landscape: who is building agent memory right now, and how the approaches differ.
- Building Memory That Scales: what breaks first as a memory system grows, and the design responses.
- The Jacobian Lens, Explained: reading a memory system through the sensitivity of its retrieval map.
- AI Memory & Context-Management Market: 2025 Update: the vendors, the categories, and where the money went.
- When AI Cites What Doesn't Exist: fabricated citations, measured, and what grounding actually costs.
Deeper reading in the Library
The longer-form clusters that grew out of this research are curated in the Library:
- Memory Science: the cognitive science and neural foundations behind machine memory, from Hopfield nets to consolidation.
- Benchmark Wars: how AI-memory systems are measured, judged, and gamed - evaluation methods and LLM-as-judge pitfalls.
- Context Builder & Orchestration: assembling the right context before the model runs - compilers, budgets, KV-cache, and MCP federation.
Featured Paper — SLoD
Semantic Level of Detail: Multi-Scale Knowledge Representation for AI Memory
Edward Izgorodin — arXiv:2603.08965
SLoD is the research operator behind Mnemoverse's multi-scale memory design. It borrows the Level of Detail concept from computer graphics and applies it to semantic memory: knowledge is represented at multiple scales, from fine-grained facts to high-level summaries, using heat kernel diffusion on hyperbolic manifolds. Production does not yet run native hyperbolic geometry — SLoD and THG are research-stage foundations; see the SLoD page for what ships versus what is proven.
Key properties:
- Hierarchical coherence: bounded error
O(sigma), distortionO(log n) - Natural abstraction detection via spectral boundary gaps
- Enables Mnemoverse to return the right granularity of memory for each query
Read the full SLoD page → | arXiv:2603.08965
Benchmarks
Mnemoverse is evaluated on standard AI memory benchmarks.
Tensor-Hyperbolic Graphs
THG page → — the graph-theoretic extension underpinning multi-hop memory association.
Looking for analysis, market landscape, and essays on AI memory? Browse the Library →.