Six systems problems define the research program.

Language models can describe a plan, simulate the voice of an expert, enumerate the steps of a build, and announce that work is complete without any necessary change occurring in the world that the language describes. This is not a minor product defect. It is a category error created by treating linguistic plausibility as operational progress.

Long-running work with language models fails less often because the model cannot generate text than because the surrounding system cannot preserve what is currently true. Transcripts accumulate proposals, corrections, reversals, sources, and outputs, but they do not reliably distinguish what was said from what was accepted. A later model must reread history and reconstruct state, often blending obsolete instructions with current intent.

Large information systems are typically navigated by names, search results, folders, timelines, and lists. These are effective but force every return to begin as a query or a reconstruction of hierarchy. Human beings also remember by place, neighborhood, landmark, scale, and relative position. FOVEA proposes a deterministic address fabric that gives digital objects stable place, procedurally generated navigational identity, and attention-proportional retrieval without turning the spatial projection into the source of truth.

Persistent AI systems need both durable truth and selective context. These requirements are often collapsed into one “memory” subsystem: a vector store, a graph, a spatial map, a long transcript, or a model-managed summary. The collapse is dangerous. The mechanism that decides what is canonical should not be the same mechanism that decides what is nearby, visible, or currently worth loading.

Conventional content-addressable storage is exact by design: identical bytes receive identical content identities; different bytes do not. Language-model systems often want a second property: repeated meanings, structures, plans, or outputs should be reusable even when their surface bytes differ. A naive “semantic hash” can create speed, but it can also erase provenance, collapse task-specific differences, cross project or authority boundaries, and promote approximate similarity into false exactness.

Most “agent” architectures begin with a running model and ask how to give it memory, personality, tools, and autonomy. That order makes identity dependent on continuous inference and encourages systems to confuse temporary computation with the person, institution, or character being represented.

Project memory in contemporary AI systems is usually implemented as transcript history, retrieved snippets, generated summaries, or opaque vendor memory. These forms preserve language but often fail to preserve operational distinctions: a proposal versus a decision, a worker return versus verification, a completed tool call versus human acceptance, an obsolete specification versus the active one.

Most multi-agent systems coordinate through conversation: one model sends another model a message, shared history grows, and completion is judged from generated reports. This is easy to prototype and difficult to trust. Inputs drift, permissions remain implicit, agents duplicate work, stale context is reused, and “done” can mean only that a model generated the word.

General-purpose compression assumes that an object is primarily a byte sequence. Many modern objects also have recoverable context: a prior version, a shared dictionary, an exact source, a deterministic generator, a model that emitted the sequence, a procedural recipe, or a lineage graph. Treating every such object as context-free bytes discards information that can be used for storage, transmission, regeneration, and random access. Treating generation as proof, however, risks silent drift and dependency laundering.

Video is usually addressed by file, URL, frame number, or timestamp. Those coordinates are useful for playback but weak for persistent meaning. A scene can move during an edit. An object can appear across shots. A generated clip can be replaced while preserving narrative role. A branch can diverge from one event and later merge. A timestamp says where bytes occur in one representation; it does not necessarily identify the enduring scene, event, object, or state.

Time in software is usually represented as an ordered sequence: log entries, frames, messages, events, commits, or samples. Sequential order is essential, but it makes multiscale reasoning expensive. The same stream must support questions about the current instant, the surrounding episode, recurring motifs, long-term drift, branch history, and exact provenance.

Transformer inference repeatedly moves and compares large amounts of state. Sparse attention, retrieval, caching, quantization, and small models reduce this burden, but their gains are often designed independently. The Glyphd fractal-computation program asks whether they can be coordinated through a shared multiscale geometry.

A language model generation already contains information that most application pipelines discard: token probabilities, partial structure, confidence changes, segment boundaries, and the moment at which a typed object becomes complete. Conventional systems wait for the entire response, then invoke a parser, classifier, second model, or postprocessor to rediscover structure that was visible during emission.

Computing devices usually require the user to name or select the object under discussion. In the physical world, however, people routinely establish reference through attention: they look at a chair, return their attention to a partner, and say “that one.” A device with camera, gaze, pose, spatial, or contextual inputs could infer candidate referents—but acting directly on inferred gaze creates the classic “Midas touch” problem, in which looking becomes accidental command.

Glyph Tokens are typed semantic carriers that compile messy human language into inspectable propositions, authorities, conditions, constraints, substitutions, uncertainties, evidence references, and actions. They are not tokenizer tokens and are not trusted merely because they are structured. Trust derives from source, authority, signature, project scope, compiler version, validators, current state, and human approval where required. The protocol is intended to make intent, routing, caching, verification, and consequential execution operate on one explicit graph while preserving exact source text and ambiguity.

What exists now: a formal architecture, phase contracts, recursive control rules, typed state handoffs, benchmark design, test schemas, and falsification criteria.