English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Lemon AI Evolving: Self-Evolving Agents That Get Smarter With Every Use

Forum topic · ✨步子哥 · 2025-11-28

Summary

This post from zhichai.net introduces Lemon AI Evolving, an AI agent framework featuring a Self-Evolving mechanism that lets agents accumulate and persist user preferences across tasks. Using a travel-planning case study, the author demonstrates how an agent learns preferences—pool-side hotels near Tokyo Tower, price ranges between $50–100, Google Maps integration, and food recommendations—during an initial 7-day Japan itinerary, then automatically applies that experience to a new US West Coast road trip task without re-prompting. The platform also offers an Agent Store where users share abstracted system-level experience (knowhow) via templates that can be replayed or remixed, solving the cold-start problem for new users. Additional covered features include a credits-based subscription model, task renaming/deletion, and task file inspection. The article argues this represents a shift from 'amnesiac' context-window-limited agents toward personalized, continuously improving AI companions, while noting open challenges around privacy and fair resource allocation.

This article introduces Lemon AI Evolving, an AI agent platform whose Self-Evolving mechanism allows agents to permanently capture user preferences in their Experience store, so they improve with continued use rather than forgetting everything between tasks.

The problem: amnesiac agents

Traditional agents remember preferences only within a single conversation. Multi-turn accumulated preferences (e.g., travel habits, brand preferences, budgets) are wiped when the context window limit is reached—LLMs can only process a limited number of tokens per session, after which earlier information is truncated.

Self-Evolving in action: the travel case study

1. First task: The user prompts for a 7-day Japan trip—vibrant neighborhoods, a hotel near Tokyo Tower with a pool, routes planned with Google Maps, and a webpage displaying results. The output is complete but lacks food recommendations and price constraints. 2. Iterative refinement: The user asks to add local restaurants and constrain hotels to $50–100 per night. The new page includes both. 3. Experience captured: In the Agent's Experience settings, preferences are persisted—pool hotels, Google Maps, food details, price ranges. 4. Cross-task transfer: A brand-new task—a 7-day US West Coast road trip in September—requires no repeated prompting. The agent automatically applies stored preferences (pool hotels, food guides, map planning, EV charging stations inferred from driving habits) to produce a tailored plan.

The essence of Self-Evolving: the more you use it, the better it performs.

Agent Store: shared intelligence

The Agent Store is a platform-level experience marketplace where user knowhow is abstracted into reusable system experience. Two ways to leverage it:

  • Automatic matching: Entering a task lets the system scan shared cases and match suitable templates (e.g., a European road trip inherits relevant system experience from Japan/US travel cases).
  • Active Remix: Browse templates, click Replay to watch the full iteration process, then Remix to reuse the experience in your own agent—an "open-source AI brain" approach that solves the cold-start problem.
  • The experience store

    The core of Self-Evolving is a dual-layer experience store: system-shared experience distilled from the Agent Store, and personal memory unique to each agent. Users can contribute their abstracted preferences back to the system, creating a positive feedback loop: more users make the system stronger.

    Credits and management

  • Subscription/credits: Upgrade via the user menu; each task consumes credits based on complexity, with transparent top-ups.
  • Task management: Rename tasks, delete tasks, and view all generated task files (webpages, maps, reports) for auditing and iteration.

Broader implications

While demonstrated with travel planning, the mechanism generalizes to business reports (remembering data styles), writing assistants (rhetorical preferences), and coding partners (debugging patterns). The Agent Store enables educators to share teaching templates and developers to remix code generators. Challenges remain around privacy (mitigated by abstraction of experiences) and fair credit allocation. Future directions may include multimodal integration (voice, video) to infer implicit preferences.

References

1. Lemon AI Evolving quick-start documentation: https://lemon-11.gitbook.io/lemon-ai-docs/lemonai-chinese/gong-neng-jie-shao/kuai-su-shang-shou 2. Lemon AI Evolving feature overview (same doc series, Agent Editor tutorial) 3. Self-Evolving agent mechanism analysis based on the Experience module 4. Agent Store sharing model (Remix/Replay mechanism) 5. AI memory persistence and LLM long-term memory optimization literature

Tags

#ai-agents#lemon-ai#self-evolving-ai#agent-memory#llm#agent-store#travel-planning#personalization

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/176415036