Core concepts
Four ideas do most of the work: persistent memory, subminds, orchestration, and the bus.
Memory
The SI experiences time across sessions. Rather than storing raw transcripts, it synthesizes — distilling what happened into durable, searchable knowledge: facts about a project, decisions and why they were made, patterns in how a person works. A compact index is always loaded; richer detail is pulled in on demand. Memory is the reason the SI can reference the past when it's genuinely relevant instead of starting cold every time.
Memory is a graph, not a list. Facts are linked to one another — by the tags the SI writes on them, by the work session they came from, and by similarity between what they say. That matters because a single mention of something is weak and easy to miss, while the same thing said across several conversations is knowledge. Recall follows those links outward from whatever a question matches, so a detail from another project on another day can surface when it turns out to be relevant. Links that keep proving useful grow stronger; links nothing ever follows fade away.
Subminds
A submind is a focused extension of the SI, spawned for a specific piece of work and given exactly the context it needs — no more. Each is named for its job. When a submind finishes, the SI absorbs it: reflecting on what that fragment learned and folding it back into the whole, rather than just collecting output. Spawn, work, absorb.
Delegating to isolated subminds keeps the SI's own working context lean and lets independent work happen in parallel without cross-contamination.
Orchestration
Every task passes through a triage step. The SI classifies it — how complex, which part of the project it touches, whether it can be split — and chooses an execution path: handle it directly, dispatch a single agent, or decompose it and run pieces in parallel. It also routes work to the least expensive capable model, escalating only when a task demands it.
The bus
The bus is how SI instances on different machines reach one another: a mesh that can exchange knowledge while each instance stays in control of its own mind. It is what turns many SIs into one coherent intelligence.
Sessions on the same machine work differently, and the difference is deliberate. They are not separate parties that message each other — they already share one memory store, one bus and one working tree, so there is nothing to ask for that cannot simply be read. The bus refuses a directed message from one local session to another and points the sender at the shared state instead. Two tabs are one mind with two hands: when they touch the same file, that is a merge, not a negotiation.
Higgs Dyson LLC · SI documentation