Eve had never seen the codebase. With Hydrate: $0.162. Without: $0.510.

Eve had never seen the codebase.

With Hydrate (Haiku): $0.162. Score: 7/10.
Without Hydrate (Sonnet): $0.510. Score: 7/10.

Same result. 3.15 times cheaper.

The cold-start problem is worst on day one. That is when the biggest model gets picked, and the whole session is spent reading files that should have been context.

Why does onboarding cost so much without memory?

A new developer with no context defaults to the biggest model (you need the reasoning headroom). Then spends half the session reading files, asking questions, establishing what the experienced team already knows.

Every one of those reads is a fresh-input token, billed at 10x the cached rate.

Memory compresses that. Not to zero. But 3.15x is not nothing.

I benchmarked new developer onboarding twice. Same codebase, same task, same automated grader. Different context.

Eve had never seen the project. In the first run, she had no injected context: cold start, no prior session memory, running on Sonnet because unfamiliar territory demands the bigger model. She explored the repository, asked architectural questions, established conventions through file reads. Session cost: $0.510. Grade: 7/10.

Second run: Hydrate injected the team's accumulated context before her first turn. Architecture decisions, established patterns, dependency policy, test conventions. She ran on Haiku. Session cost: $0.162. Grade: 7/10.

Same grade. 3.15 times cheaper.

The reason is straightforward once you see it. A new developer without context has two costs: the model upgrade (Haiku cannot handle true cold discovery as reliably as Sonnet) and the discovery tax (reads, questions, establishing what the team already knows). Both are priced at fresh-input rates, which are ten times the cached rate.

Memory removes the second cost almost entirely and makes the first cost unnecessary. The injected context block is stable, served from cache from turn one, and tells Eve what would otherwise require a full repository tour to discover.

This is the onboarding argument for team memory. Not that new developers make fewer mistakes. That they spend less of their first session re-learning what the team already documented.

Full benchmark data: gethydrate.dev/blog/memory-not-smarter-cheaper-models

Eve's onboarding benchmark. Never seen the codebase.

Cold start on Sonnet: $0.510, 7/10.
Warm context on Haiku: $0.162, 7/10.

Same grade. 3.15x cheaper. The cold-start problem hits hardest on day one.