AI Agent Memory Costs Explode Without Validated Compaction

A new paper splits AI agent memory into five stages, per Gaurav Dadhich's arxiv paper. Only one stage, done right, keeps costs from spiraling.

Teams running agents in production eat that cost. Accumulate history naively and it grows quadratically as conversations get longer. Summarize it crudely and the curve flattens back to linear, but at an accuracy cliff, where real information gets cut along with the noise. A reference build that gets this right, called Maximem Synap, still scores 92% on LongMemEval and 93.2% on LoCoMo.

Dadhich names the five stages as architecting, ingesting, scoping, anticipating, and compacting and consolidation. The last one is where the money is: only compaction validated against accuracy loss gets both linear cost and preserved fidelity. The framework treats memory as a resource shared across an organization's agents and teams, not a single user's chat log.

Those scores measure recall. They don't cover the things the paper says no benchmark captures: latency, token efficiency, and resistance to context rot. That gap means 92% and 93.2% describe how well the system remembers, not how fast or cheap it runs.

Teams buying bigger context windows to fix agent memory are solving the wrong problem. The paper's cost curve says compaction quality decides the bill, not window size, and nobody outside Dadhich's paper has benchmarked that yet.