Computation is killing collaboration

Computation is killing collaboration.

The biggest hurdle in AI-assisted development isn’t a lack of computation - it’s our love for it.

When engineers hit a “memory” issue with an LLM, the instinct is to treat it as a technical challenge alone:

• Throw more tokens at the window
• Spin up another vector database
• Build infrastructure for that infrastructure
• Design agent flowcharts that look squigglier than my attempt at Arabic calligraphy

We are building massive “recursive” scaffolds when what we actually need is a simple prosthesis.

In my recent work on Simulated Recall via Shallow Indexing (SR-SI), I’ve found that the context problem is architectural, not computational.

The LLM doesn’t need to remember your entire 100k LOC codebase in its neural weights. It just needs to know where to look, and to look before it acts.

By externalizing project knowledge into a self-managed, markdown-based shallow index, I’ve seen:

• ~106x cumulative improvement in token coherence
• An 85.5% reduction in context-maintenance costs
• Zero re-teaching overhead across thousands of prompts

The folly: we treat memory like a storage problem. In reality, it’s a retrieval strategy problem.

As context windows get bigger, the linear scaling of traditional workflows will only lead to more cognitive mirage - the illusion that the AI is following you when it’s actually just drifting.

SR-SI doesn’t compete with bigger windows; it gets exponentially stronger because of them. It turns the AI from a stateless assistant into a persistent system collaborator using nothing more than Git and a bit of structural discipline.

Stop building squiggly overcomplicated flowcharts and infrastructure, and focus on building cognitive scaffolding.

As a follow up to my whitepaper, I will soon be releasing a working plan to explain how SR-SI can work at scale using two simple solutions that anyone can implement using: .md files and Git.

SR-SI: The methodology that gives AI persistent memory across any long-running project