OpenMAIC v1.3.0: Tsinghua Puts the Classroom into a Knowledge Base — the Agent Remembers Every File You Upload
Upload a 50-page slide deck to an AI assistant, ask how the formula on page 3 is derived — it answers. Close the browser. The next day, ask how that formula relates to the figure on page 15, and it says: "What formula?"
This is the common flaw of all "stateless" AI education tools — every conversation starts from zero, materials must be re-uploaded, and context must be rebuilt.
OpenMAIC from Tsinghua's MAIC lab released v1.3.0 on 2026-10-29, and the core update is one sentence: every upload enters the knowledge base, is kept permanently, and is available across conversations.
Knowledge Base: From "One-off Materials" to "Persistent Memory"
Before v1.3.0, material handling in OpenMAIC was per-session — you upload a PDF, the agent parses it, you can ask questions within that conversation. When the conversation ends, the material is gone.
What changed in v1.3.0:
1. Persistent storage: every upload is saved in the owner's knowledge base until actively deleted 2. Folder organization: materials can be organized into folders instead of one flat pile 3. Parse once, use forever: materials are parsed once at upload; all subsequent conversations can reference them 4. @ references: mention materials in the knowledge base with @ in the composer; the agent will read, search, and organize them
What does this resemble? A teacher's filing cabinet. Materials used in each class are stored there, not thrown away when class ends. Next time you prepare a lesson or answer questions, you pull directly from the cabinet.
Pro Workbench: Preview First, Then Generate
The other v1.3.0 update is the Pro workbench. The core change: request previews.
Previous flow: give an instruction → agent generates the full lesson → you review → revise if unsatisfied.
Pro workbench flow: give an instruction → agent first shows a preview ("here's what I plan to do") → you confirm or adjust → only then does the agent formally generate.
This is the application of "judgment-gate decoupling" in education: generation and judgment are separated. The agent first presents a plan (judgment); execution (gate) happens only after you confirm. This avoids the waste of "finding out the direction was wrong after generation is done."
Multi-Agent Classroom: Not One AI Teaching You, but a Team Teaching a Class
OpenMAIC's core architecture is multi-agent collaboration:
- Classroom chat agent: answers student questions about any slide element, interactive content, or whiteboard drawing
- Generation agent: generates lesson content, slides, and video
- Editing agent: "Edit with AI" directly edits slides
- Search agent: web research capability introduced in v1.2.1, searching the web via Kimi/DeepSeek/Firecrawl
- PPTX: PowerPoint files preserving each scene
- Offline HTML: a single HTML file that plays the entire lesson
- Frontend: TypeScript / Next.js
- Backend: server-first architecture (stable since v1.2.0), requires PostgreSQL
- Deployment: Vercel deployments stay at 1.1.x (server-first needs a long-running server)
- Models: built-in Pi model; also supports Kimi, DeepSeek, Firecrawl, and other providers
- License: MIT open source
- GitHub: https://github.com/THU-MAIC/OpenMAIC
- License: MIT
- Language: TypeScript
- Stars today: 115
- Latest release: v1.3.0 (2026-10-29)
- Source: Tsinghua University MAIC lab
These agents share the same knowledge base and lesson context. When a student asks a question, the chat agent can call the search agent to query the web, cite materials from the knowledge base, and give a sourced answer.
One-Click Export: Classrooms Become Shareable Assets
v1.2.0 introduced classroom export:
This means an AI-generated lesson can be used independently of the OpenMAIC platform — sent to colleagues, uploaded to an LMS, or embedded in a webpage.
Tech Stack and Deployment
Security Track Record
OpenMAIC's update history shows serious security practice:
| Version | Date | Security Fix | |---------|------|--------------| | v1.0.1 | 09-06 | Hardened default configuration | | v1.0.2 | 09-14 | SSRF and DNS rebinding fixes | | v1.0.3 | 09-15 | Access-code token expiry, Next.js RCE patch | | v1.1.1 | 09-27 | MinerU Cloud parsing hardening | | v1.1.2 | 09-28 | Provider requests only connect to verified addresses, reject redirects |
Four of five releases contained security fixes, indicating repeated real-world testing and hardening.
Why This Matters
Competition among AI education tools has moved from "can it generate courseware" to "can it remember the student."
The first generation of AI education tools (2023–2024) solved generation: given a topic, produce a deck. The second generation (2025–2026) solved interaction: students can ask questions and agents answer. The third generation is solving memory: the agent remembers what the student has learned, what they don't understand, and what materials they need.
OpenMAIC v1.3.0's knowledge base is a concrete implementation of the third generation — not memory in model parameters (which fades), but externally stored memory (precise, auditable, organizable).
This forms an interesting contrast with "parameters as storage" (SKIM): model parameters can store factual knowledge (closed-form writes into FFN weights), but education needs structured, retrievable, organizable materials — better suited to an external knowledge base.
Data
OpenMAIC is not the hottest project (115 stars/day), but it represents an important direction: AI education tools are moving from "one-shot generation" to "persistent teaching companions." The knowledge base gives agents cross-conversation memory, and the Pro workbench adds a judgment gate before generation. Together, these two updates mean the relationship between student and agent shifts from "one lesson" to "one semester."