AI Pulse · Under the Radar

A four-prompt technique for building persistent agent memory is showing up in multiple places at once, and it matters because it sidesteps the usual machine learning stack entirely.

The method, credited to Boris Cherny (who built Claude Code), chains four Claude prompts together to extract entities, resolve duplicates, build relationships, and store everything in a knowledge graph. No training, no embeddings, just structured prompts feeding into each other. The same week the technique surfaced in a thread on X, at least two independent repos shipped nearly identical pipelines. One of them, Graphify, hit 848 stars per day, making it one of the fastest-growing repos in the current wave.

Why this matters if you run a company: most AI memory systems today rely on vector search, which is good at fuzzy matching but terrible at tracking relationships over time. A customer who bought three things, returned one, and escalated twice is not just a vector. A knowledge graph holds those connections explicitly, so your agent can reason about history, not just similarity. The Cherny approach makes that buildable without a data science team.

What changes

If this pattern holds, you will see more agents that remember context across sessions without needing a dedicated ML pipeline. That means faster iteration, fewer dependencies, and memory that you can inspect and edit like a database. It also means the barrier to building stateful agents just dropped. The kind of persistent context that used to require a research team is now four prompts and a graph store. This is exactly the sort of capability we wire into the Omni Command Centre, where memory and reasoning need to scale across departments without adding headcount.

The repos are young, the technique is unproven at scale, but the convergence is real. If three builders land on the same design in the same week, the design is probably right.