Controlled comparisons
Memory approaches are compared with baseline and cold-start controls to examine where prior experience helps.
Research
ALATAR
Memory and coordination for sustained AI work.
ALATAR brings together persistent-memory research and the engineering of systems that coordinate AI across coding tools.
Internal researchVallinor AI · Toronto, Canada
01 / Persistent memory
VARIE
Persistent memory. Sustained mathematical research.
Research aim
Persistent memory, put to the test.
VARIE investigates how AI can retain useful experience and carry it into extended research. The program spans combinatorial search, analytic number theory, geometric optimization and quantum many-body physics, combining controlled memory experiments with internal computational verification.
Computational verification
Mathematical candidates are checked internally with exact computation and, in selected investigations, separate implementations.
Investigation domains
Research across mathematics and physics.
From covering designs to quantum spin systems, the work spans constructive search, certified numerical studies and exhaustive finite checks.
Analytic number theory
Riemann zeta function and the behavior of L-functions near the critical line.
Quantum many-body physics
Energy bounds and spectral-gap questions in quantum lattice systems.
Combinatorics
Covering designs and Ramsey-type questions about unavoidable patterns.
Geometric packing
Circle-packing questions and the efficient arrangement of geometric objects.
Research chronology
From memory engineering to mathematical investigation.
Building the memory foundation
Initial engineering established persistent research memory, source tracking and retrieval, followed by staged testing of how earlier work is carried into subsequent tasks.
82 covering-design entries improved
Research used the La Jolla covering-design archive as a reference for continued search. Covering designs ask how to cover every required combination using a compact collection of overlapping groups.
La Jolla / Campaign results82covering-design entries improvedAcross multiple research runs, the campaign improved 82 distinct entries relative to a frozen archived baseline.
All 15 constructions in a selected verification set passed independently implemented checks.
Campaign total across multiple runs and experimental conditions. Improvements were verified internally. External confirmation remains open.
Observed follow-up comparison17 paired rounds with 45-minute budgets60Managed memoryimprovement events28Persistent notebookimprovement eventsBoth approaches retained memory. An event is a verified improvement over two frozen archived baselines and that approach’s earlier result; successive improvements to the same problem can each count. This was an exploratory comparison in one domain using one model.
Mathematical investigations across four domains
The research portfolio includes constructive combinatorics, internally certified numerical studies, geometric optimization and finite quantum-system calculations. Selected investigations combine exact checking with separate internal implementations.
Geometric optimization / Circle packing15 research episodes · August–September 2026One sustained internal run improved the arrangement of equal-sized circles inside a larger circle across 42 distinct cases.
42verified circle-packing improvements10h 30mmetered active run timeImprovements were checked against fixed reference packings under the run’s declared criteria. The time figure covers this run; development, other campaigns and separate rechecks are outside it. External confirmation remains open.
Selected research highlightsConstructive combinatorics A checked construction in Ramsey theory
Investigating how large a network can remain without prescribed patterns becoming unavoidable—a lower-bound question in classical Ramsey theory.
Internal validationA candidate construction passed two exact internal checks and a separate reimplementation. External confirmation remains open.
Geometric combinatorics Exhaustive exclusion in a finite cover model
Studying which local arrangements can occur in a finite geometric cover model.
Internal validationExhaustive checks excluded tested patterns within a defined small class, supported by separate internal reconstruction and review. The broader question remains open.
Analytic number theory Numerical studies of zeta and L-function zeros
Investigating the spacing of zeta-function zeros and the distribution of L-function zeros near the critical line.
Internal validationInternally certified numerical work, with separate checks and assumptions recorded for each result.
Quantum many-body physics Testing quantum energy-gap criteria
Investigating whether proposed criteria can establish an energy gap in kagome-lattice quantum spin models.
Internal validationSelected finite-system calculations passed separate internal checks; conclusions are limited to the models and criteria studied.
An internally verified covering-design improvement with Claude Opus 5
A research run using Claude Opus 5 produced an internally verified improvement over an archived reference value for a covering-design problem. External confirmation remains a separate milestone.
02 / Coordination
MANWE
Coordination for complex agent work.
Research aim
One project. Multiple coding tools. Coordinated work.
MANWE grows out of Overseer, a desktop research prototype for coordinating AI coding agents across tools. The engineering problem is to turn a project brief into parallel work while keeping responsibilities, progress and human direction clear.
Research prototypeCross-tool coordination
Prototype integration work covered Cursor Agent, Claude Code and Grok, exploring coordination across their command-line coding environments.
Parallel work with clear ownership
The design organizes a project into assigned workstreams and isolated task workspaces, with planning, implementation and review treated as separate responsibilities.
Human direction during a run
The prototype engineering explores progress tracking, pause and resume controls, and a feedback path for steering work as requirements change.
Named tools and models identify research and prototype environments. Their inclusion does not imply affiliation with or endorsement by their providers.
Discuss MANWE