Let an AI agent turn every meeting recording into a knowledge base you can ask

An indie developer who ships AI in real projects keeps a recording running in meetings and everyday chats. Feishu turns each finished recording into a transcript automatically. Every day, his AI agent pulls the new transcripts into his knowledge base, writes a summary, and files the people, projects and to-dos separately. By his own public account, he has built up more than 170 transcripts in a few months. Now he can just ask, "Who did I talk to this week, and what still needs follow-up?" What makes it get smoother with use is that every recording gets broken into people, projects and to-dos. The more people the AI recognizes, the better it files the next recording.

The recording knowledge base at a glance

He only does one thing: he turns on the recorder during meetings. An AI agent handles every other step on a schedule:

Stage	What happens	Output
Capture	A recorder or meeting app records and syncs on its own, and generates a transcript	One full transcript
Pull	The AI agent pulls new transcripts into the knowledge base every day	The original is saved and never changed after that
Summarize	It reads the whole transcript and writes notes: what was said, what was decided, and who owns which to-dos	A one-page summary of the meeting or conversation
File	It writes people, companies, projects and to-dos onto their own pages	People pages, project pages, a to-do list
Ask	"What did I focus on this week?" "Who should I team up with on this?"	Answers with sources, plus one weekly report

Over time, each meeting links up with the people in it and the projects it touches. Open a person or a project and you can see every conversation in date order.

The numbers above are the author's own public claims. We haven't verified the 170-plus transcripts, and we haven't run this setup ourselves. He uses Anker's Soundcore recorder with Feishu (Lark), which lets the AI agent pull new recordings on a schedule. You don't need Feishu, though. Any tool that exports a transcript works: meeting transcripts from Zoom, Google Meet or Teams, or recorders like Otter and Plaud. Drop the exported transcripts into one inbox folder and have your AI agent process new files there every day. Everything after that works the same.

Why it works: three habits

• Originals are stored unchanged: the full transcript is saved as is, and the summary goes on a separate page. The AI's answers always point back to the exact words, so a bad summary can't mislead you about what was really said.
• Knowing people matters more than remembering events: once a person has a profile, the next time they show up in a transcript the AI can recognize who you're talking to from what they say. New faces get checked with you first. The AI doesn't guess.
• You only press record: pulling and summarizing run on a schedule, and so does filing. You don't have to remember to export and tidy up, so it keeps going for months without gaps.

Build it: a five-step SOP

Hand this to your AI agent. After each step, have it check its work against the "done when" column. When a step is finished, it should stop and wait for your OK. The author's open-source knowledge base foundation is already packaged as a set of skills. See the jinchenma-knowledge-base Skill.

Safety basics

• Tell people before you record: before you start, tell everyone in the room what the recording is for. If someone says no, don't record. Many places have clear laws on recording, and a team should have a written rule.
• Keep sensitive topics out: don't record salary, HR or private client matters. Before you share a summary with anyone, remove names and sensitive details.
• Split access by department and project: before you start, decide who can read which meetings, and whether the knowledge base lives on your computer, a private server or the cloud.
• Read-only access, and no credentials in chat: the recording platform only gets read access. The transcription service key stays in a local file and never goes into the chat.

Common pitfalls

• Keeping only the summary, not the original: a summary is the AI's interpretation, and it can be wrong. Without the original, you can't check it when you find a mistake.
• Letting AI create profiles for new people on its own: similar names and accents can make it pick the wrong person, and transcription errors can too. One wrong profile can misfile every recording after it, so always confirm new faces first.
• Rolling it out to the whole company on day one: let yourself or one small team use it for a month or two first. Once the rules run smoothly, expand. Otherwise, access and consent get messy fast.
• Ignoring transcription quota and cost: once the recording platform's quota runs out, transcripts can get cut off. If you switch to a paid transcription service, estimate the monthly cost first.
• Using it commercially without checking the license: this knowledge base foundation is CC BY-NC 4.0, which allows non-commercial use only. Before a company uses it officially, contact the author for permission.

Project: jinchenma-knowledge-base (CC BY-NC 4.0, non-commercial use only). Want a simple local knowledge base that anyone can use first? Read the tutorial Turn Obsidian into a personal knowledge base AI can search and write to.