Why multi-agent multiplies the problem
One stateless agent already forgets context; running several in parallel multiplies it. Each has its own window, and without coordination they make conflicting assumptions and decisions.
The challenge shifts from 'remember context' to 'share consistent context across agents'.
One shared source of truth
All agents should read the same project memory file for conventions, decisions, and constraints. A single source of truth keeps them aligned instead of each inventing its own rules.
If agents read different (or no) memory, divergence is guaranteed.
Scope each agent
Give each agent a clear, bounded slice — a module, a feature, a layer — so their work doesn't overlap. Narrow scope also keeps each agent's context focused and accurate.
Clear ownership prevents two agents 'fixing' the same thing in incompatible ways.
Isolate parallel work
Run parallel agents on separate branches or git worktrees so simultaneous edits don't collide, then integrate deliberately with review.
Isolation turns 'conflicting edits' into 'normal merges'.
Re-sync between tasks
When one agent makes a decision that affects others, write it to the shared memory and have the others re-read before continuing.
Keeping the shared memory current is what keeps a fleet of agents coherent.