Agent Bottlenecks Aren't in Context, but in Memory's Write, Change, Recall, Forget
Uber burning through its token budget in 13 weeks proves stuffing context is wrong; MongoDB points to memory as the performance bottleneck—write, change, recall, forget—and suggests starting with employee use cases and human-in-the-loop.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
Database schemas follow the scarcest resource
Pete starts with E.F. Codd's relational model paper from 1970: when storage was the most expensive thing, normalized tables were the rational choice. By 2007, when MongoDB was born, the scarce resource had become time, so the document model deliberately denormalizes, using JSON/BSON to reduce disk reads. He also corrects a repeatedly misread positioning: MongoDB is not 'schemaless' but 'schema-flexible'—the same collection can hold documents of different shapes, letting the schema evolve with the product. This design logic shows that database schemas should follow whatever resource is scarcest at the time.
— Pete JohnsonStuffing context with data degrades quality
The news that Uber burned through its 2026 token budget in 13 weeks is treated as a sample of token maxing strategy failure. Stuffing large amounts of data into context is only half about cost; quality also drops: research shows the first and last few thousand tokens of a window matter most, the middle is mostly noise. This path pushes the industry toward 'classified memory'—selecting only the few terms relevant to the current agent loop each round, rather than dumping all history back. The key here isn't saving money but the attention decay inside the context, making selective retrieval a performance issue.
— Nathan LabenzAI features grow from where customers already hurt
MongoDB's search capabilities grew layer by layer from customer pain points: lexical search first in 2020 because many customers were building their own Lucene; then vector search because vectors are essentially float arrays and can naturally be a document field; then combining lexical, vector, and metadata pre-filtering into one hybrid search. In 2025 they went further and acquired Voyage AI, bringing embedding models in-house for a 'better together' offering. This sequence itself shows: infrastructure companies touching AI features usually start from where customers already hurt, not by buying models first and then inventing use cases.
— Pete JohnsonChunk size tuning should be delegated to the model
The first nagging problem in RAG is chunk size: chunks too small mean insufficient context and poor retrieval; too large inflate storage and hurt precision, forcing manual trial and error. Voyage's 'contextual chunking' sends the sentence and extra context as two separate strings into the model, which automatically decides the optimal chunk size and returns a single vector. This achieves retrieval quality close to full context with very small chunks, at lower storage cost, and is now in its fourth version. It's a concrete example of moving the engineer's iterative tuning process into the model's weights.
— Pete JohnsonEmbedding models aren't commoditized; selection drives hallucinations
Pete gives two selection thresholds: on scale, you really start to feel pain around 100k vectors; on quality, Voyage leads on Hugging Face's RTEB benchmark, improving retrieval by up to 14% over other embedding models—a gap that can directly show up as hallucinations. He also reminds that embedding models are not commoditized—Anthropic doesn't make its own embedding model, so his advice is to use Voyage directly. For many teams, embedding model selection has long been underestimated; this segment puts it back at the center of performance and hallucinations.
— 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, but agents for only about eighteen months. Long-term context strategy is no longer about stuffing the whole conversation back; it's about 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 phrase: Write, change, recall, forget. Memory has a half-life—recent information matters more than information from weeks or months ago. This means agent memory must be treated as a first-class citizen, with intentional writing, updating, and retiring.
— Pete JohnsonFirst enterprise AI wave lands on employees, not customers
Most Fortune 500 companies Pete talks to are working on employee-facing use cases and keeping humans in the loop. Two reasons: ROI is easier to calculate because these roles already have KPIs; and data security consequences are lighter—employee data leaks are easier to handle than customer data leaks. Conversely, fully autonomous agents that face customers are a worse risk-reward bet right now. This segment provides a coordinate for enterprise AI budget decisions: the first wave is more likely in internal efficiency tools than in robots making final decisions for customers.
— Pete JohnsonMost sophisticated customers aren't in the US but in Mexico City and São Paulo
Pete met this year's most sophisticated customers in Mexico City and São Paulo, and these customers typically assume US competitors are more advanced. He says what he sees is the opposite. He attributes this to the proliferation of hyperscale data centers and AI services—geographic barriers matter less than in the cloud era, leveling the ability of global enterprises to access technology. This observation suggests that country labels are increasingly unreliable for judging whether a company can use cutting-edge AI.
— 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 share of database market | about 3% | 12:24 |
| Time for Uber to burn through 2026 token budget | 13 weeks | 16:38 |
| Voyage contextual chunking versions | third version released last summer, fourth released in the last six weeks | 40:10 |
| Voyage retrieval quality improvement over other embedding models | up to 14% | 48:20 |
| Retrieval quality improvement from using a reranker | 5-10% | 1:03:40 |
| MongoDB acquisition price for Voyage AI | $220 million | 1:17:00 |
| Voyage acquisition price as share of MongoDB market cap | less than 1% | 1:17:20 |
Glossary
- token maxing
- Stuffing as many tokens as the budget allows into context; costly and quality degrades in the middle.
- contextual chunking
- Sending a sentence and extra context as two inputs into a model, which automatically decides the optimal chunk size and returns a single vector.
- Matryoshka embedding
- An embedding model that supports truncating high-dimensional vectors to lower dimensions, avoiding re-embedding the entire corpus.
- agent skills
- Curated system prompts from MongoDB, covering topics like data modeling and operations, as markdown files.
How to listen
Engineers building Agent/RAG applications, technical leads choosing data stacks for enterprises, investors tracking embedding models and vector databases.
The first 30 minutes lean product-promotional; start at 16:38; around 31:51 is a feature release, can fast-forward.