Build Your Personal Knowledge Base With AI (Step by Step)

Summary

Nick Saraev argues that a personal knowledge base can make AI responses more reliable by giving a model searchable access to your own information. Instead of sending a large archive to the model every time, the system searches stored data for relevant material and supplies only the best matches.

He contrasts this approach with putting everything directly into a model’s context window. Large context windows can hold substantial amounts of material, but too much context may increase cost and latency, reduce retrieval accuracy, and cause context degradation. Older and newer facts may also be treated as equally relevant. Saraev proposes dating incoming information and giving newer material greater relevance.

Saraev says his company’s system, which he calls Engram, holds roughly 35,000–50,000 entries, including Git history, company Claude sessions, and video transcripts. When asked a question, an AI selects search terms, retrieves relevant records, and uses them to answer.

Practical Guide: Building the Knowledge Base

1. Identify information sources. Examples include email, website and lead records, Slack messages, phone-call transcripts, and AI chat sessions.
2. Create connectors to collect new data. Each connector should retrieve only information added since its previous run and attach a date and time. Saraev’s Gmail connector uses read-only access. He says the initial sync collected a year of email, while later runs captured only new messages. He schedules the connector every 15 minutes, though the interval can vary.
3. Index AI conversations. Saraev recommends preserving ChatGPT or Claude sessions, which he says are stored locally as plain text and can otherwise be deleted after 30 days. An archive can help recover past decisions and avoid repeating work.
4. Generate searchable metadata. For each incoming item, use an AI model to create a short summary, notes describing its contents, and likely search questions. Also extract precise “atoms,” such as error messages, fixes, working commands, and reasons behind decisions. This can turn implicit organizational knowledge into explicit institutional memory.
5. Use straightforward text search. Saraev says a conventional search algorithm, BM25, can be sufficient; vector embeddings and vector databases are not always necessary. BM25 ranks documents based on search-term frequency. The AI can then consult the original source for full context.
6. Keep the pipeline running. Monitor connectors and fix failures caused by changed API keys, permissions, or routes. Without regular updates, the knowledge base becomes stale.

Saraev says his implementation uses an inexpensive model to produce summaries and metadata, and that the system returns results quickly. He also claims plain-text search substantially reduced inference time and maintenance costs compared with approaches he had previously built using RAG and vector databases.

Key Points and Caveats

• The system is only as useful as the data it receives and the reliability of its connectors.
• Date-based relevance weighting can help prevent obsolete information, such as old product prices, from being favored.
• The method supplements AI’s memory by searching an external archive; it does not require the model to reread the entire archive for every question.
• Saraev describes the process as quick to set up, but the full system involves connectors, metadata generation, search, and ongoing monitoring.

Reviews, Guides, or Tutorials

• Guide/tutorial: A step-by-step explanation of creating an AI-searchable personal or company knowledge base.
• Comparison/analysis: Saraev says his tests found this approach more effective than earlier RAG systems he had tried. He argues that plain-text BM25 search can outperform more elaborate setups for this use case.
• No standalone product review is presented.

Main Speakers or Sources

• Nick Saraev — The video’s narrator and the source of the system design and performance claims.
• Vannevar Bush — Cited for his 1945 vision of a machine that could serve as an extension of memory.
• Cohere — Cited as the inspiration for the data-upload pipeline Saraev adapted.
• Boris Cherny — Cited by Saraev as having praised regular search relative to other approaches.