Skip to main content

How a piece of work flows through Memex AI

A Strategy is created

Someone identifies a problem worth solving or an objective worth pursuing. They create a Strategy, not a one-line epic title, but the full context: what’s the problem, who does it affect, what does the solution look like architecturally, what principles should guide the work. This is the document that every decision and every work item will trace back to. Without it, agents are building features. With it, they’re solving a problem.

Non-obvious design choices surface

As the strategy takes shape, Decisions are logged within it. Each has a stable ID, options with trade-offs, and a status. Some are resolved quickly; others require research, prototyping, or stakeholder alignment. The strategy gives each decision its context. D7: Cache invalidation approach is meaningless alone, but within the strategy it’s clear why this choice matters and what’s at stake.

Work is scoped

Work Items are defined within the strategy, with goals, acceptance criteria, and explicit dependencies on other work items AND specific decisions. The dependency graph makes it clear what can start now and what’s blocked. Every work item inherits the strategy’s context, an agent loading WI-3 can always navigate up to the strategy to understand the broader objective.

Decisions are resolved

Through research, discussion, prototyping, or stakeholder input, decisions get resolved. Each resolution records the choice, the rationale, and what was rejected. Resolved decisions unblock work items and update affected blueprints.

An agent picks up a work item

The agent connects via MCP and:
  • Reads the work item spec (goal, dependencies, checklist)
  • Loads all relevant blueprints (deployment, conventions, architecture)
  • Checks that all blocking decisions are resolved
  • Produces an execution plan reconciling the spec against the actual codebase
  • Waits for plan review before writing code

Implementation happens

The agent executes the plan. If it discovers the code contradicts a blueprint, it flags drift. If a new design question arises, it creates a new decision. When complete, it updates the work item status.

Blueprints update

Resolved decisions and completed work items trigger blueprint reviews. Blueprints are updated to reflect the current state of the system. Drift detection runs continuously.

The cycle continues

New decisions surface. New work items are scoped. Blueprints evolve. The graph grows, but it stays current because the system enforces it.