How Xiaohongshu Built an Enterprise AI Personal Assistant from Zero to Full‑Staff Coverage

Timeline and AI‑Native Approach

The team decided on February 11 to build an AI personal assistant. A three‑person group delivered an internal‑test version in three days. By March 9 the architecture was upgraded and 2,000 public‑test slots opened. By March 17, after two weeks of scaling, the assistant covered all employees – roughly one month from decision to full rollout.

AI‑Native Project Mechanism

The core was an "AI Native" workflow: define a new mode, ship an MVP quickly, validate with real users, then scale. Early user feedback drove rapid iteration, avoiding the traditional 90%‑ready evaluation that would make AI solutions obsolete before deployment.

Security Isolation – Seal AI Zone and NEX Sandbox

Security was addressed by creating a dedicated Kubernetes cluster called Seal AI Zone , physically isolated from production. Employees access it via a Seal client. Inside, a multi‑layer security stack includes:

MicroVM‑based sandbox ( NEX – Native Engine for Agent Sandbox) with six isolation layers (VM boundary, Seccomp whitelist, default‑deny network, read‑only filesystem, cgroup limits, full audit logs).

Cold‑start P50 <300 ms , hot‑start P99 <50 ms , hot‑pool hit rate >70%, creation throughput peak 500 ops/s.

Data Privacy

All user input passes a Security SDK that performs three scans (PII, secret keys, file content). If no sensitive data is found, the request goes to external LLM APIs; otherwise it is routed to a private, on‑premise model, ensuring that sensitive data never leaves the corporate network.

Cost Optimization

Cost was tackled on two orthogonal dimensions: token consumption and model price.

Self‑GC (Token‑Saving)

Inspired by JVM garbage collection, Self‑GC treats the long‑term dialogue history as a virtual heap. It assigns identities to text fragments, then asynchronously plans which parts to fold, mask, or prune. The three‑phase commit (Async Plan → Rehearsal → Delayed Commit) ensures no abrupt context loss.

In production, Self‑GC reduced average token usage by 10‑15% during the day and up to 20% at peak, yielding substantial cost savings at enterprise scale.

Auto Model Routing (SealRouter)

SealRouter decides whether a request needs a high‑end model or a cheaper one. It combines a classification model trained on live traffic with rule‑based overrides (e.g., many tools or long context trigger a higher‑tier model). Cache‑aware cost constraints avoid unnecessary model switches.

Routing results: ~75% of requests use cost‑effective models, 22% use a higher‑tier model, and only 3% invoke the top‑tier model, cutting overall model cost by 69% without sacrificing quality.

Three‑Layer Memory Architecture

To overcome limitations of the existing memory system, a new three‑layer design was introduced:

L0 Session – short‑term context, transient state, not persisted to long‑term memory.

L1 Work Log – fact buffer with evidence; flushes to dated markdown files, preserving raw source pointers.

L2 LLM Wiki – structured, compiled markdown wiki serving as long‑term knowledge. Unlike traditional RAG, facts are pre‑compiled and retrieved on demand via BM25 + vector search, enabling continuity, traceability, filtering, and maintainability.

Ecosystem – Skill Hub and Cowork Studio

Skills are abstracted as reusable agents on Skill Hub , supporting official, personal, and private‑domain contributions. Increased usage fuels more skill contributions, creating a positive feedback loop.

Cowork Studio provides a one‑click publishing platform for internal AI applications, lowering distribution friction.

Future Directions for Enterprise Agents

The team envisions the next generation of agents evolving along three axes:

Active perception – agents proactively gather external information and discover opportunities.

Goal‑driven behavior – agents understand high‑level objectives, decompose them into tasks, and execute plans autonomously.

Collaborative agents – agents act as independent members within an organization, coordinating with other agents and humans.

Overall, using an AI‑Native methodology, a three‑person team launched the assistant in three days and achieved full‑staff coverage in a month, solving security and cost challenges while building a scalable, extensible AI ecosystem.