The Agent Bottleneck Isn't Context, It's Memory: Write, Change, Recall, Forget
Uber burning through its token budget in 13 weeks proves stuffing context is the wrong move; MongoDB points the Agent performance bottleneck at memory — write, change, recall, forget — and suggests starting with employee use cases and humans in the loop.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
Database schemas follow whatever resource is scarcest
Pete takes E.F. Codd's 1970 relational model paper as the starting point: storage was the most expensive thing then, so normalized tables were the rational choice. When MongoDB was born in 2007, the scarce resource had become time, so the document model deliberately denormalizes, using JSON/BSON to cut disk reads. He also corrects a positioning that gets misread over and over: MongoDB is not ‘schemaless’ but ‘schema-flexible’ — the same collection can hold documents of different shapes, letting schema evolve along with the product. The design logic says: a database schema should follow whatever resource is scarcest at the time.
— Pete JohnsonFilling context with data actually lowers quality
The news that Uber burned through its 2026 token budget in 13 weeks is treated as a sample of token maxing strategy failing. Cost is only half the problem with stuffing large amounts of data into context; quality also drops: research shows the first and last few thousand tokens of a window matter most, and the middle is basically noise. This path pushes the industry toward ‘categorical memory’ — each turn selects only the few terms relevant to the current agent loop, rather than dumping all history back in. The key here isn't saving money, it's that attention decay inside the context makes selective retrieval a performance problem.
— Nathan LabenzAI features should grow from where customers already hurt
MongoDB's search capabilities grew layer by layer around customer pain points: lexical search first in 2020, because many customers were building their own Lucene; then vector search, because a vector is essentially a float array and can naturally be a document field; then lexical, vector and metadata pre-filtering into the same hybrid search. In 2025 it simply acquired Voyage AI, bringing the embedding model in-house to form a ‘better together’ offering. The order itself says something: when infrastructure companies touch AI features, they usually start from where customers already hurt, not by buying a model first and then inventing a scenario.
— Pete JohnsonChunk size should be left to the model, not human trial and error
The first grinding problem in RAG is chunk size: too small and there isn't enough context, so retrieval quality suffers; too large and storage grows while precision actually drops, leaving only manual trial and error. Voyage's ‘contextual chunking’ sends the sentence and extra context into the model as two separate strings, lets the model automatically determine the optimal chunk size, and returns just one vector. This gets retrieval quality close to full context using very small chunks, with lower storage cost, and is now on its fourth version. It's a concrete example of moving the engineer's repeated parameter tuning into model weights.
— Pete JohnsonEmbedding models aren't commoditized; the choice decides hallucination
Pete gives two selection thresholds: on scale, vector volume only really starts to hurt around 100,000; on quality, Voyage leads on the Hugging Face RTEB benchmark, improving up to 14% over other embedding models, a gap that may show up directly as whether there are hallucinations. He also warns that embedding models have not been commoditized — Anthropic doesn't build an embedding model itself, so his advice is to just use Voyage. For many teams, embedding model selection has long been underrated; this segment puts it back at the center of performance and hallucination.
— Pete JohnsonMemory has a half-life and must be actively retired
Why is memory hard? Pete's analogy: the industry has been building databases for sixty years, and doing agents for about eighteen months. The long-context strategy is no longer stuffing the whole session back in, but recalling the best context at query time and writing new facts back to memory after answering; enterprise systems also use RBAC to control memory sharing and isolation. He sums up the whole loop with a colleague's words: Write, change, recall, forget. Memory has a half-life, and recent information matters more than information from weeks or even months ago. This means agent memory must be treated as a first-class citizen, with deliberate writes, updates and retirement.
— Pete JohnsonThe first wave of enterprise AI lands on employees, not customers
Most Fortune 500 companies Pete deals with are doing employee-facing use cases, and keeping humans in the loop. There are two layers to why: ROI is easier to calculate, since these roles already have KPIs; and the consequences of a data security breach are lighter, since employee data leaks are easier to handle than customer data leaks. Conversely, fully autonomous customer-facing Agents are the bet with the worse risk-reward ratio right now. This segment gives a coordinate for enterprise AI budget judgment: the first wave is more likely to land in internal efficiency tools, not robots making final decisions for customers.
— Pete JohnsonThe most complex customers aren't in the US, they're in Mexico City and São Paulo
Pete met his most complex customers this year in Mexico City and São Paulo, and these customers usually assume US competitors are more advanced. He says what he saw was the opposite. The reason is attributed to the spread of hyperscale data centers and AI services: geographic barriers matter less than in the cloud era, and the ability of global enterprises to access technology has been leveled. This observation suggests that judging whether a company can use frontier AI makes country labels less and less reliable.
— Pete JohnsonIn their own words · checked verbatim
Most people think that we're schemaless, and that's not actually true. It's not that we're schemaless, it's that we're schema flexible
Pete Johnson23:21
There's no lamp stack for agents yet in the way that we have with web development. We will get there. Having lived through that having lived through that life cycle, we'll eventually get there. But, like, there's there's no React in Angular. There's no LAMP stack for agents right now.
Pete Johnson1:10:04
Most of Fortune five hundreds that I talked to are Dewey employee facing use cases with human in the loop for exactly the reasons that you just laid out and for some of the ROI reasons that I laid out before.
Pete Johnson1:13:10
there's a presumption in other countries that The US is ahead and doing things that other people are not, and I found the opposite to be true.
Pete Johnson1:24:24
every other embedding model is gonna force you to pay tokens during your development cycle to hit against their embedding model.
Pete Johnson1:28:08
Figures
| MongoDB's database market share | about 3% | 12:24 |
| Time for Uber to burn through its 2026 token budget | 13 weeks | 16:38 |
| Voyage contextual chunking versions | the third version shipped last summer, the fourth shipped in the last six weeks | 40:10 |
| Voyage's retrieval quality improvement over other embedding models | up to 14% | 48:20 |
| Retrieval quality improvement from using a reranker | 5-10% | 1:03:40 |
| Price MongoDB paid to acquire Voyage AI | $220 million | 1:17:00 |
| Voyage acquisition price as a share of MongoDB's market cap | less than 1% | 1:17:20 |
Glossary
- token maxing
- The practice of stuffing every token in the budget into context, which is costly and degrades quality in the middle.
- contextual chunking
- Sending the sentence and extra context into the model as two inputs, automatically determining the optimal chunk size and returning a single vector.
- Matryoshka embedding
- An embedding model that supports truncating high-dimensional vectors to lower dimensions, avoiding the need to re-embed the entire corpus.
- agent skills
- Curated system prompts MongoDB provides, as markdown files covering topics like data modeling and operations.
How to listen
Engineers building Agent/RAG applications, technical leads choosing data stacks for enterprises, investors watching the embedding model and vector database space.
The first 30 minutes lean toward product promotion; you can start at 16:38; around 31:51 is a feature launch and can be fast-forwarded.