Service & Retail AI Watch, Issue 1: 2B Agents and Memory Capability

Summary AI initiatives in service retail are diversifying across both 2B and 2C fronts: on the 2B side, merchant “digital employees” and AI business-development agents; on the 2C side, consumer-facing service-retail agents. Across the board, these projects track the broader industry’s stage of development — early, with product-market fit still being validated for some and technology-product fit still being explored for others.

Against that backdrop, tracking AI in service retail needs to systematically watch both which scenarios are landing and how the underlying technology is evolving — useful as an external reference for current projects and as groundwork for how the business responds to the AI shift. This issue covers the business-application side through the commercially faster-moving “2B agent,” and the implementation side through a core building block of any agent: memory.

---

Key Findings

Business Application — 2B Agents

• 2B agents are landing in a “cost reduction → efficiency gain → decision-making” progression. Customer service and sales scenarios have clear ROI and high standardization, making them the preferred beachhead; marketing and operations carry more upstream/downstream complexity, and adoption is picking up as model comprehension and generation capability improve; operating-strategy scenarios demand the most from both model capability and merchant trust, and are the hardest to land (both Alibaba’s 1688 and ByteDance have pushed diagnostic-type features to the back of their roadmaps).
• 2B agents land in different industries on a gradient — the pattern is “e-commerce → advertising → local services,” starting wherever digital infrastructure and merchant sophistication are strongest.
• Workflows and agents tend to coexist as a “guardrail + delight” pairing. Workflows handle stable execution of structured tasks and keep outcomes predictable (e.g. standardized image generation); agents are dynamically driven by the underlying model and suit open-ended Q&A, delivering a stronger “sense of a person” and more flexibility.

Implementation Framework — Memory Capability

• Different customer segments want fundamentally different things from memory:
  • Long-tail B2B merchants: prefer “invisible memory” that lowers the cognitive load of understanding the system (e.g. 1688 never surfaces memory settings at all).
  • Key-account merchants: need deep, controllable management to support complex customization (e.g. Coze’s structured memory features).
  • Consumer users: are comfortable with contextual memory and respond well to a simple, transparent management layer (e.g. ChatGPT lets users view and edit what it remembers).

---

I. Introduction

1.1 AI Industry Moves

2025 is the pivotal year AI agents move from tooling assistance toward industrialized deployment, as breakthroughs in function calling, the MCP protocol, reasoning models, and reinforcement fine-tuning (RFT) drive a paradigm shift from “tool-assisted” toward “autonomous decision-making.”

• Early March: The world’s first general-purpose agent, Manus, launched — capable of independently thinking, planning, and executing complex tasks and delivering finished output end to end, setting a new state-of-the-art score on the GAIA benchmark. A milestone moment for the agent field.
• Mid-May: Flowith’s NEO Agent impressed, posting a GAIA score above both Manus and OpenAI’s Deep Research. It brought several breakthroughs: (1) uninterrupted 24/7 cloud task execution; (2) no practical limit on input/output token length; (3) meaningful gains across every GAIA difficulty tier (L1–L3).
• Late May: At Google I/O 2025, Google upgraded Gemini and shipped a suite of tools. In service-retail-adjacent areas, AI Mode and Gemini Live showed real deployable capability: (1) one-tap ticket and restaurant booking inside AI Mode; (2) virtual try-on in apparel e-commerce via uploaded photos; (3) Gemini Live executing phone actions from voice commands, combined with visual recognition for tasks like appliance installation or bike repair.

1.2 Service Retail Industry Tracking Plan

(Tracking roadmap omitted here — see internal business analysis framework.)

---

II. Introducing Agents

“AI Agent” is typically translated as an intelligent entity capable of perceiving its environment, making decisions, and executing actions.

Large language models (LLMs), thanks to their planning and reasoning abilities, were quickly adopted as the substrate for agent construction. The architecture most widely accepted in the industry is the one proposed by Lilian Weng, OpenAI’s Head of Applied AI Research:

Agent = LLM + Planning + Memory + Tool Use

On the front end, agents present in two branches (which can converge):

1. Code-invocation style (Devin, Genspark): the operating process is abstracted into an input-output black box — well suited to fintech and other contexts that must interface with core systems and prevent screen-recording leaks.
2. Human-simulation style (UiPath, YingDao, Playwright): drives the browser to mimic human keyboard/mouse actions. Simulating human behavior is more vivid and intuitive than code, and builds trust between people and the agent faster.

At the architecture layer, Anthropic drew an important distinction:

• Workflows: systems that orchestrate the LLM and tools through predefined code paths. Human-driven, with the advantage of being stable and transparent — well suited to repetitive, predictable tasks.
• Agents: the LLM dynamically drives its own process and tool use. The advantage is strong generalization, able to handle unfamiliar environmental information — well suited to open-ended, dynamically changing problems.

---

III. Service Retail AI Benchmarking

2C agents are mainly segmented by industry; 2B agents mainly by functional domain. The core requirement on the consumer side is delight; on the business side, it’s a specialized knowledge base or vertical training, where stability of problem-solving matters more.

3.1 Business Application — 2B Agents

This section focuses on the commercially more mature 2B agent space. Trends:

• Domestic vs. overseas: domestically, incumbent platform players dominate (built around their own merchant base); without a Databricks-equivalent platform, breaking down knowledge silos within an enterprise is hard. Overseas, startups penetrate more easily because Western enterprises already completed digital transformation in the SaaS era.
• Evolution path: AI-native companies (e.g. 11x) focus on a single scenario (e.g. SDR); non-AI-native companies (SaaS, CRM, RPA incumbents) are more inclined to build a multi-industry, cross-scenario, full-chain agent capability.
• Application scenarios: sales agents (11x, Salesforce, etc.) are currently the leading revenue-generating scenario, since the traditional cost of acquiring a Sales Qualified Lead (SQL) is high.

1) Goal Design As of mid-2025, most 2B agents still set goals at the tool level, not yet tied to business-throughput metrics. Evaluation approaches fall into two camps:

• User-behavior-centric: adoption/conversion-rate style metrics. E.g. 1688’s digital employee tracks click-through on recommended topics; ByteDance’s AI customer service uses transaction conversion as its metric.
• No mature framework yet: e.g. Xiaoshouyi (Kingdee’s CRM) is still exploratory, with some teams falling back on release-iteration counts or complaint-resolution rates.

2) Agent Landscape Comparison

Platform / Product	Company Type	Scale Signal	Agent Type	User Base	Service Domain	Core Scenario / Highlight	Core Users	Key Success Factors	Base Model
1688 Digital Employee	Domestic platform	Substantial daily active usage	Workflow + Agents	Sizeable daily active base	Marketing, ops, business management, customer service	Digital employees (business manager, product/customer/promotion specialists)	1688’s long-tail merchants plus some KA customization	1) Merchants already have data accumulated 2) User-friendly operation 3) Rich APIs and vertical models	Qwen (primary)
ByteDance	Domestic platform	Sizeable enterprise beta cohort	Agents-first	—	Marketing, ops, business management, customer service	Livestream/short-video AIGC features	E-commerce merchants and agencies with operating capability	1) Content-platform asset accumulation	DeepSeek, in-house commerce/vision models
Coze	Domestic platform	—	Workflow-first	—	—	—	—	—	Doubao (self-developed)
Xiaoshouyi (CRM)	Domestic platform	Large SaaS customer base	Workflow + Agents	Small trial cohort	Sales	Marketing assistant, sales assistant, sales coach	Large MNCs with SaaS experience	1) Customer-resource moat 2) Tencent ecosystem advantage 3) Feature-rich	Third-party integration (no fine-tuning)
Salesforce	Overseas platform	Very large market cap, huge SaaS customer base	Workflow-first	Large user base, sizeable paid cohort	Sales, customer service	Custom workflows, lead management, sales coaching	Mid-to-large enterprise customers across industries	1) Customer-resource moat 2) Mature commercialization 3) Deep and broad PaaS layer	GenCode, ChatGPT, Gemini
YingDao RPA + AI	Domestic startup	Meaningful valuation	Agents-first	Unknown	Sales, ops, business management, customer service	External data scraping/analysis, store-ops automation	Mostly e-commerce customers	1) Low delivery cost 2) Deep e-commerce case-study base 3) Active community	Third-party integration (no fine-tuning)
11x	Overseas startup	Meaningful valuation	Workflow + Agents	Unknown	Sales	Lead sourcing/screening, personalized email, outbound calling	SMBs across industries	1) Fast iteration 2) Task-based pricing 3) High per-step efficiency	Third-party integration

3.2 Implementation Framework — Memory Capability

1) Definition and Importance Agent memory is the mechanism for storing, organizing, and managing interaction information. It is, in essence, an extension of the LLM’s limited context — enabling personalized interaction, keeping context coherent, and letting the agent plan future actions based on historical data.

2) Implementation Paths Two mainstream approaches exist today:

• Attention-mechanism-based: focuses on the model’s internal mechanics, implemented from the training side.
• RAG-based (Retrieval-Augmented Generation): persists interaction information to external storage. E.g. Mem0 delegates the key memory operations to the LLM itself, implementing memory reasoning and updates through prompt engineering.

3) Memory Taxonomy

Human Memory	Agent Memory	Example
Sensory memory (lasts seconds)	Raw input (text, image, or other modality)	Seeing an image, your first impression of it
Short-term memory (working memory, ~20–30s)	Input context, bounded by the transformer’s context window	The few digits you hold in your head during mental math
Long-term memory (unbounded, explicit/implicit)	External vector store, unlimited capacity, fast runtime retrieval	Riding a bike — a skill you still have years later

4) Agent Memory Landscape Scan Scanning leading agents, user-interaction design for memory falls into three patterns:

1. Invisible memory: e.g. 1688 — no management surface exposed, to lower the comprehension cost for B2B users.
2. Deep management: e.g. Coze — the most complete memory management built for KA B2B customers.
3. Simple management: e.g. ChatGPT — high acceptance among consumer users, with lightweight management exposed.

Modality Capability / Product	ByteDance	Coze	Dify	ChatGPT	1688 Digital Employee
Text memory	✅	✅	✅	✅	✅
Multimodal memory	✅	✅	❌	💡 (in progress)	❌
Enable/disable control	✅	✅	✅	❌	❌
Memory editing	✅	✅	✅	❌	❌
Turn-level management	✅	✅	❌	❌	❌
User presets	✅ (variables, database, file box)	✅ (preferences, conversation history, etc.)	❌	❌	❌

---

IV. Appendix

A1. Evolution of AI Agent Technology

Before the LLM era, agent technology went through several stages:

• Symbolic agents: relied on symbolic logic to encode knowledge, excelling at explicit reasoning frameworks (e.g. expert systems), but struggling with large-scale uncertainty.
• Reactive agents: emphasized fast, real-time response, mapping input directly to output with no higher-level planning.
• RL-based agents: learned by trial and error against an environment to maximize reward (e.g. AlphaGo). Advantage: autonomous learning; challenge: long training time, low sample efficiency.
• Transfer- and meta-learning agents: transfer learning accelerates adaptation to new tasks; meta-learning (“learning to learn”) helps infer an optimal policy from very few examples.

A2. Alibaba 1688 Digital Employee

Built a “virtual employee” persona on top of the underlying AI operations capability, to make module-by-module management easier.

• Employee types: Business Manager, Product Specialist, Customer Specialist, Promotion Specialist.
• Dialogue vs. skills: Q&A leans toward consultation; skills are executed automatically via the agent. Common query types are “how do I do X,” “what’s my number/report,” and casual chat.
• Operating flow example (“set a promotional price”):
  1. Merchant selects a strategy and clicks “View.”
  2. AI estimates impact using category and data signals, and recommends a discount plan.
  3. Merchant clicks “Apply now.”
  4. AI executes with one click; merchant reviews the result.

A3. ByteDance and Coze

ByteDance tailors its approach by business line:

• E-commerce: encourages merchants to self-design and invoke flows on Coze.
• Local services: the platform provides free AI customer-service configuration for single stores (beta stage), aiming to raise the floor for small merchants.
• Advertising: offers pre-built industry agents (currently covering education, photography, beauty, and legal services).
• Scenario priority: AIGC (livestream/short-video) > AI customer service > compliance scenarios > comparison monitoring > summary/analysis.

A4. Xiaoshouyi (CRM)

Still positioned primarily as a tooling platform; its core edge is integration with existing SaaS customer data. At the model layer, it uses an open-source SOTA model without fine-tuning. The team favors agile development and releases agents (marketing, analyst agents, etc.) cautiously and incrementally.

A5. YingDao RPA + AI

A software-robot development platform with a meaningful valuation.

• Main products: YingDao RPA (traditional office automation), YingDao AI Power (RPA + AI, covering an intelligent assistant, Chat-BI, etc.), Data Connect, Low-Code, and more.
• Core scenarios: external data scraping/analysis (e.g. competitor monitoring), store-ops automation (e.g. bulk listing generation, auto campaign submission).
• Key success factors: efficient delivery (“teach a person to fish” model), a deep base of e-commerce case studies, and a very active developer community.

A6. Salesforce

Evolution path: from assistant (Einstein Copilot) to agent (Agentforce).

• Einstein Copilot: primarily user-guided, responding to queries.
• Agentforce (AI Agent): executes tasks autonomously, goal-driven, handles complex dynamic scenarios, and reduces hallucination via the Atlas reasoning engine and RAG.

A7. 11x

A London-based AI sales startup with a meaningful valuation. Its flagship products are digital workers Alice and Julian (mapped to SDR and AIBD roles respectively).

• Insight: can’t yet fully replace a senior salesperson, but delivers strong value for cost. It raises automation by chaining together the individual steps of the SDR workflow.
• Trend: sales agents are gradually consolidating coverage across the full sales lifecycle; pricing is shifting from “per seat” toward “pay for outcome.”
• Moves: introducing voice-to-voice models; rebuilding its system to integrate a range of tools (integration completed within hours).

A8. 1688 Digital Employee Memory Capability

• Memory tiers: real-time memory (VRAM) → short-term memory (Redis, on the order of a few months) → mid/long-term memory (RAG vector database).
• Evaluation framework: focuses on latency, accuracy, contextual consistency, satisfaction, and coverage.
• Limitations and roadmap: frequent strategy changes create “noise interference” that disrupts memory; the roadmap calls for a multimodal image-text product feature library to improve retrieval success.

A9. Glossary

• SDR (Sales Development Representative): handles lead capture, initial outreach, and needs analysis.
• Function Calling: the technique by which an LLM calls external functions/tools (a calculator, a database query) via an API to execute a specific task.
• MCP (Model Context Protocol): an open protocol, introduced by Anthropic, that standardizes how LLM applications connect to external tools, data sources, and systems.
• RFT (Reinforcement Fine-Tuning): reinforcement learning used to improve a model’s performance on a specific task.
• GAIA difficulty tiers (L1–L3):
  • L1 (basic): simple information retrieval and processing tasks.
  • L2 (intermediate): computation tasks that involve, e.g., spreadsheet files.
  • L3 (advanced): complex reasoning and multi-tool sequencing tasks.
• Playwright / UiPath: Playwright is a leading browser-automation tool; UiPath is a global leader in traditional RPA.
• Anthropic: an AI-safety-focused company founded by former OpenAI executives, maker of the Claude model family.