Do I need a vector database for a personal knowledge base?

Not always. Keyword search with BM25 over well-written summaries works well when your notes use consistent terms, and it is cheaper and faster. Semantic or hybrid search helps when the same idea is written in many different ways.

Is BM25 search a form of RAG?

Yes. RAG means retrieving relevant material and adding it to the model's prompt before it answers. BM25 is one way to do the retrieval step, and embeddings with vector search are another.

What is recency weighting in AI retrieval?

Recency weighting stamps every record with a date and boosts newer records in search results while lowering older ones. It helps when facts change, such as prices. It can also hide old information that is still true, so some systems mark records as replaced instead.

All video notes

Video notes · Notes and PKM

The NEW Way to Build a Second Brain w/ Claude (2027)

Video by Nick Saraev on YouTube · 12 min · uploaded October 2026

Notes and analysis by BrainTube, written from the video. Not affiliated with Nick Saraev. Watch on YouTube

In short

Nick Saraev describes the company knowledge base he calls Engram: connectors pull new data every 15 minutes, a cheap model writes summaries and "atoms" for each item, everything is date-stamped and recency-weighted, and Claude searches it with BM25. The pipeline is practical and mostly right. His claim that this means you do not need RAG is a labelling mistake, since keyword retrieval feeding a prompt is still RAG, and BM25 alone has blind spots his summary step only partly covers.

The argument

Saraev argues that large context windows do not remove the need for a knowledge base. Dumping everything into the window is costly and slow, it suffers from context rot, and the model treats a two-year-old price the same as today's. The fix is a lean retrieval pipeline. Define your sources, write small connectors that fetch only new items and timestamp them, have the cheapest trusted model write a summary, likely search questions and a list of atoms (exact errors, fixes, commands, decisions) for each item, then search those summaries with plain BM25 and boost recent results. Keep the connectors healthy, because stale data breaks the whole thing.

Key ideas, with timestamps

1:01
Why large context windows are not enough
He accepts that current windows can hold roughly 35 to 50 books' worth of text, then gives three reasons not to rely on them: cost, time, and accuracy. His example is an old product price in the knowledge base being treated as equal to the current one when drafting a proposal.
2:33
Recency weighting
Every item gets a date stamp when it enters the system. Recent items get a boost at search time and old ones a penalty. He pairs this with periodic sweeps that remove items no longer worth keeping.
3:34
Engram at company scale
His system holds about 35,000 to 50,000 records: git history, the company's Claude sessions and his video transcripts. When he asks Claude a question, a fast step picks a search term, results come back in about half a second, and the best matches are added to the prompt.
5:08
Connectors that fetch only what is new
Sources might include email, the website, lead conversations, Slack and VoIP call transcripts. His Gmail connector uses read-only API access, pulled a year of mail from three accounts on the first run, and now runs every 15 minutes. He stresses read-only access as a safety rule.
6:11
Index your AI sessions
He recommends adding Claude sessions because they are already plain text on your machine and are normally deleted after 30 days. Indexing them stops you repeating yourself and keeps decisions from earlier projects.
7:14
Summaries, likely questions and atoms
A cheap model, Claude Haiku in his case, writes a short summary for each item, including the question someone would type to find it. It also extracts atoms: exact error messages, working fixes, commands and the reasons behind decisions. Search runs on the summaries, which point back to the full raw item.
9:17
BM25 instead of vectors
He says dropping vector databases cut inference time and embedding costs by about nine tenths. BM25, a word-frequency ranking method from the 1990s, does the matching and Claude infers the rest. He cites Boris Cherny of Claude Code as finding plain search beat other approaches.
10:19
Keep the pipes fresh
Rotated API keys, changed permissions and broken routes will silently stop connectors. He recommends a background check that reruns and verifies each connector, because a knowledge base that stops updating is worse than none.

What it gets right

Ingestion is the hard part. Saying that connectors matter most is correct and underrated. Most failed knowledge bases fail because data stops arriving, not because the search is weak. Incremental fetches, timestamps, read-only scopes and a health check are the right defaults.

Write for retrieval, not storage. Generating a summary plus the questions someone would ask is a known technique for improving search over raw text. Extracting atoms, such as exact error strings and the reason for a decision, is the most useful idea in the video. Those are the details people search for later, and they are often buried in long threads.

Date everything. Stale facts are a real failure mode, and a date on every record is the minimum needed to deal with them.

Where it is wrong or thin

This is still RAG. He presents BM25 as a way to avoid RAG, but searching a store and adding the results to the prompt is retrieval-augmented generation, whatever the ranking method. What he dropped is embeddings and a vector database. That distinction matters, because it keeps the useful lesson clear: lexical search is often enough, and cheaper.

BM25 has blind spots. Keyword ranking misses paraphrases and synonyms. Searching "cancel subscription" will not reliably find a note that says "churned". His question-writing step helps, and Claude can retry with other terms, but for notes written in varied language a hybrid of keyword and semantic search usually finds more. The Claude Code finding he cites came from searching code, where exact names matter most.

Recency is a blunt rule. A decaying score helps with prices, but it also buries old facts that are still true, such as a founding decision or a contract term. Marking records as replaced when newer ones arrive is more precise than weighting by age.

Not every chat is on your disk. Local plain-text transcripts with a 30-day default cleanup describe Claude Code (as of the October 2026 upload). Chats in the ChatGPT or Claude web and desktop apps are stored by the provider, so they need an export rather than a file indexer.

Privacy is skipped. Pushing every email, Slack message and call transcript through a model for summaries raises access questions inside a company. Who can search the CEO's inbox? He does not say. The closing claim that multi-billion dollar companies run the same system is not backed by anything in the video.

How this applies to memory across AI tools

His Engram is built for one company and one assistant. Most individuals want the same pattern without writing connectors: their own sources, searchable from whichever AI tool they use that day.

For reading and watching, BrainTube covers that ground: it saves YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base and exposes it as an MCP server, so Claude, ChatGPT, Cursor and other MCP clients can search it. You still need your own pipeline for company systems like email and Slack.

For the retrieval background, see RAG explained and semantic vs keyword search.

Who should watch it

Watch it if you are building a knowledge base for a team or business and want a lean architecture that skips vector databases. The whole video is about twelve minutes; the most useful parts are 5:08 to 7:14 (connectors and session indexing) and 7:14 to 10:19 (summaries, atoms and BM25). Skip the opening on context windows if you already know about context rot. Skip the video if you want a step-by-step build: there is no code on screen, and the full prompt sits behind a free sign-up linked in the description.

Frequently asked

Do I need a vector database for a personal knowledge base?: Not always. Keyword search with BM25 over well-written summaries works well when your notes use consistent terms, and it is cheaper and faster. Semantic or hybrid search helps when the same idea is written in many different ways.
Is BM25 search a form of RAG?: Yes. RAG means retrieving relevant material and adding it to the model's prompt before it answers. BM25 is one way to do the retrieval step, and embeddings with vector search are another.
What is recency weighting in AI retrieval?: Recency weighting stamps every record with a date and boosts newer records in search results while lowering older ones. It helps when facts change, such as prices. It can also hide old information that is still true, so some systems mark records as replaced instead.

Keep videos like this one searchable

BrainTube saves YouTube videos with their transcripts into one compiled knowledge base that Claude, ChatGPT, Cursor and other MCP clients can search.

Start free

The NEW Way to Build a Second Brain w/ Claude (2027)

In short

The argument

Key ideas, with timestamps

Why large context windows are not enough

Recency weighting

Engram at company scale

Connectors that fetch only what is new

Index your AI sessions

Summaries, likely questions and atoms

BM25 instead of vectors

Keep the pipes fresh