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

Forum topic · ✨步子哥 · 2025-11-28

Summary

Traditional AI agents suffer from a form of amnesia: preferences shared in one conversation, such as travel style, budget limits, or brand loyalty, vanish in the next task because of context window limits. Lemon AI Evolving addresses this with a Self-Evolving mechanism that permanently stores user preferences in each agent's Experience library. In the platform's travel-planning demo, a 7-day Japan itinerary is iteratively refined (adding restaurants and a $50-100 hotel budget), and those captured preferences automatically carry over into a later US West Coast road-trip task with no re-prompting. Beyond personal memory, the Agent Store lets users share abstracted knowhow: new agents can auto-match relevant templates, or users can Replay and Remix community cases, solving the cold-start problem. A credit-based subscription model funds task execution, while task management features (renaming, deletion, file viewing) keep agent workspaces organized. The forum post frames this as a shift from AI as a disposable tool to AI as a learning companion that improves through use and collective experience.

Introduction

Most AI agents today behave like forgetful assistants: they can perfectly follow your preferences within a single conversation, but forget everything in the next task. This forum post introduces Lemon AI Evolving, a platform built around a Self-Evolving concept where agents accumulate experience and genuinely improve with use.

Why Traditional Agents Forget

Conventional agents lose multi-turn preferences because of the context window limit: an LLM can only process a limited number of tokens per session (thousands to tens of thousands). Once exceeded, earlier information is truncated, causing preferences to evaporate, much like human short-term memory holding only 7±2 items before needing conversion into long-term memory.

Tell a typical agent your detailed preferences once (traveling with dogs, EV charging stations, Marriott hotels under a budget, cultural over natural sights) and it delivers a perfect plan. Start a new task, however, and everything resets to zero.

The Self-Evolving Mechanism

Lemon AI Evolving permanently captures user preferences in each agent's Experience library:

1. First task: Create a new agent with a prompt like: "I want a 7-day trip to Japan. I like vibrant neighborhoods, want a hotel near Tokyo Tower with a pool, plan routes with Google Maps, and build a webpage showing all results." The agent produces a polished itinerary with routes and hotels, but misses food and pricing details.

2. Iterative refinement: The user adds: "Include local restaurant recommendations and keep hotel prices between $50 and $100." The updated page now includes food picks and budget-locked hotels.

3. Memory capture: In Agents settings > Experience, the recorded preferences (pool hotels, Google Maps, food details, price range) persist as cross-task permanent experience.

4. New task, inherited wisdom: Requesting a 7-day US West Coast road trip for September requires no repetition. The agent automatically injects prior learnings, pool hotels, food guides, map-based planning, even EV charging stops inferred from driving habits, producing a tailored travel guide.

The core principle: the more the agent is used, the better it performs.

Agent Store: Shared Knowhow

The Agent Store turns individual experience into a platform-level knowledge marketplace:

Automatic matching: When creating an agent, the system scans shared cases and applies matching templates. Planning a "European road trip" recalls system experience from similar Japan/US travel cases.
Replay & Remix: Browse the template store, Replay a case to watch the full iteration process (prompt tuning, preference injection), then Remix it to inherit that experience into your own agent, an "open-source AI brain" approach.

Shared knowhow is abstracted into reusable system modules (e.g., "travel planning always includes map integration + price filtering"), raising first-attempt success rates and solving the cold-start problem.

The Experience Library

The experience library has two layers:

System-shared experience: distilled from Agent Store contributions
Personal memories: each agent's unique user habits and preferences

It forms a positive feedback loop: personal preferences can be abstracted and uploaded to benefit others, while users absorb collective wisdom, so more users make the system stronger.

Credits and Task Management

Subscription: Click username > Upgrade > pay to obtain credits. Each task consumes credits based on complexity; recharge from the credits management page.
Rename tasks: Task list > More > Rename, to keep workspaces organized.
Delete tasks: Same path, one click to clean up.
View task files: Click a task to see all generated files (webpages, maps, reports) for auditing and iteration.

Applications and Outlook

The mechanism generalizes beyond travel: business report agents that remember your data style, writing assistants that retain rhetorical preferences, coding partners that accumulate debugging patterns. Challenges remain around privacy (addressed via experience abstraction/anonymization) and credit fairness (transparent subscriptions). Future directions may include multimodal integration so agents can infer implicit preferences from voice and video.

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