mem·o·ry sprawl
noun/ˈmɛm.ə.ri sprɔːl/
The uncontrolled accumulation of persistent AI-agent memories across tools, vendors, and sessions — without shared governance, lifecycle management, or any reliable way to know what your agents remember about you.
The architectural fragmentation of agent memory itself: the same facts duplicated across vector stores, graph databases, and document stores, each with its own permissions, none with a plan for forgetting.
"Legal asked which systems hold customer memories. Nobody could answer. Classic memory sprawl."
"We have eleven agents and nine memory stores. The sprawl audit starts Monday."
Etymology
- shadow IT
- SaaS sprawl
- agent sprawl
- context sprawl
- memory sprawl ← you are here
Why it's about to matter
In 2025, agents got memory. Every major platform shipped some version of it: persistent stores that let an assistant remember your preferences, your projects, your colleagues, your last mistake. Memory is what turns a chatbot into a coworker, so everyone built it at once.
Nobody shipped forgetting.
The result is already visible in any organization running more than a handful of agents: memories accumulate per-agent and per-vendor, duplicated and contradictory, with no shared inventory, no retention policy, and no answer to the simplest questions — what do our agents know, where is it stored, who can see it, and how do we delete it? Enterprises spent a decade learning to govern data at rest. Agent memory is data that acts, and it is sprawling faster than any data class before it.
Sprawl always gets named after it's too late to prevent and just in time to sell the fix. This page exists to name it slightly earlier than that.