Assistant Mode¶
Assistant mode prioritizes conversation continuity, personalization, and source-linked compaction. It enables conversation, knowledge, persona, entity, short-term, and summary memory by default.
Build an assistant¶
from memorizz import ApplicationMode, MemAgentBuilder
agent = (
MemAgentBuilder()
.with_name("Onboarding assistant")
.with_instruction("Help users onboard and distinguish facts from guesses.")
.with_application_mode(ApplicationMode.ASSISTANT)
.with_llm_config({"provider": "openai", "model": "gpt-4o-mini"})
.with_memory_ids("onboarding")
.with_semantic_cache(enabled=True, threshold=0.88)
.build_and_save()
)
answer = agent.run(
"Remember that I prefer dark mode.",
memory_id="onboarding",
user_id="user-42",
thread_id="first-run",
)
Design guidance¶
- Put stable user or organization facts in entity memory; keep one-off turn details in conversation memory.
- Keep persona changes versioned and auditable. A transient user instruction should not rewrite the agent's durable identity.
- Set semantic-cache freshness by domain, and invalidate cached policy or catalog answers when the source changes.
- Generate summaries with explicit memory, user, and thread scope.
- Add external tools only with typed schemas and appropriate approval policy.
Assistant mode is a good starting point for support, onboarding, and internal help desks. Use workflow mode when repeatable tool execution matters more than conversational history.