Harrison Chase, co-founder and chief executive of LangChain, says memory is rare in AI agents today and still early. “One thing about agents that is rare today but will be incredibly common in the next few years is this memory layer,” he said on the 20VC podcast, in a clip Harry Stebbings posted on X on 17 October. “I think memory is super, super early.”
In a second clip, posted early on 18 October, Chase said consumers will use both general and specialised agents, “and the horizontal agents will interact with the vertical agents.” LangChain says it works with 35% of the Fortune 500, and its Series B announcement says it raised $125M at a $1.25B valuation.
Business Pill · AGENT MEMORY
A short explainer on agent memory: how an AI assistant keeps what it learns about you between conversations. It teaches the general idea only and says nothing about LangChain or Harrison Chase.
The Full 20VC Episode
Harry Stebbings’s conversation with Harrison Chase on 20VC, published on YouTube on 17 October 2026.
The key insight: As we read it, Chase is placing the open question in agents above the model: he calls the memory layer “rare today” and “super, super early,” while naming Instinct and Muse as products that have “nailed personalization and learning.”
Chase named two consumer agents as his examples: “If you look at Instinct or Muse, one of the things that they have nailed is personalization and learning, making it seem like it is something for you.”
Stebbings also posted written notes from the conversation, which he says Chase allowed him to share. They summarise the point this way: Instinct and Muse “have nailed personalization and learning, so the product feels like it is for the individual user, through its tone or proactive follow-ups.”
The full 20VC episode, published on YouTube on 17 October, runs just over an hour. Its chapter list includes “Why the Agent Memory Layer Is Still Wide Open” and “Will Every Consumer App Have Its Own Vertical Agent?”
LangChain says it works with 35% of the Fortune 500, has crossed 1 billion open source downloads and ingests over 1 billion events per day on LangSmith; its Series B announcement says it raised $125M at a $1.25B valuation. Source: LangChain About page and Series B announcement.
What Was Posted on X
Why the memory layer will become more and more important and how Instinct crushed it
โOne thing about agents that is rare today but will be incredibly common in the next few years is this memory layer.
If you look at Instinct or Muse, one of the things that they have nailed is personalization and learning, making it seem like it is something for you.
I think memory is super, super early.โ @hwchase17
Love to hear your thoughts @DhravyaShah @taranjeetio @noahrshinn @natfriedman
Harry Stebbings, host of 20VC, posted this 20-second clip on 17 October in which LangChain CEO Harrison Chase calls the agent memory layer rare today and super, super early, naming Instinct and Muse as examples. It is the source of the quotes on memory above.
Will we have one horizontal agent or tons of vertical agents in the consumer realm?
โMy take on the consumer space is that you will have both horizontal and vertical agents, and the horizontal agents will interact with the vertical agents.
A lot of these apps have knowledge about you and can learn things about how you interact with that app better than a general horizontal agent could.
It is a better experience if there is an agent that wraps those APIs and exposes a natural language face for a horizontal agent to interact with.โ @hwchase17
Love to hear your thoughts @jerryjliu0 @Altimor @garrytan
In this clip, posted on 18 October, Chase says consumers will have both horizontal and vertical agents, with horizontal agents interacting with vertical ones that wrap each app’s APIs. It is the source of the quote card and the horizontal-and-vertical section.
SGLang is now named in more AI job ads than LangChain.
That one comparison says a lot about where AI hiring is heading.
In the job ads we track: (1) Inference engines – vLLM: 138 ads at 47 companies – SGLang: 107 – TensorRT: 93 (2) Agent frameworks – LangChain: 99 – LlamaIndex: 16
Three things to read into these findings: 1. The money moved, and the jobs followed. Gartner expects inference to make up 55% of AI-optimized cloud spending in 2026, the first year it passes training. Once the biggest line item is running models rather than building them, someone has to own cost per token, latency and GPU utilization. Companies now name the tools for that job right in the ad. 2. Agents created the bill. An agent that plans, calls tools and retries makes many model calls to finish one task. Frameworks decide which calls happen. Inference engines decide what each call costs. The more agents ship, the more the second skill is worth. 3. Look at who is hiring. The most vLLM ads come from Nebius (22), Baseten (9) and FuriosaAI (9). One is a cloud that sells open models by the token. One is an inference platform whose revenue run-rate tripled to $600M in a single quarter. One is a Korean chip startup. None of them is building an app. All of them sell compute, and how efficiently they serve it sets their margin.
FuriosaAI is the most telling. It pitches its serving software as a drop-in replacement for vLLM.
For a new AI chip, raw speed is no longer enough as it has to run the engine developers already use. vLLM is turning into the layer that models, clouds and chips all have to speak.
Investors see it too. vLLM’s creators raised $150M for Inferact at an $800M valuation in January. SGLang’s team raised $100M for RadixArk at $400M in May.
For engineers, making each call fast and cheap is a solid career now.
On 18 October @agentnativedev posted counts from the AI job ads it tracks: vLLM named in 138 ads at 47 companies, SGLang 107, TensorRT 93, against LangChain 99 and LlamaIndex 16. The counts are the poster’s own tracking, not figures we checked; we include it because it puts Chase’s framework, LangChain, next to the inference engines that hiring now names.
// Latent Semantic Memory for LLMs //
This paper presents some really interesting findings about agent memory.
They find that sixteen memory tokens beat full-context inference on FactKG while using 108x fewer tokens for the source documents.
Approach:
Documents are encoded once into reusable embeddings. For each query, a learned module gathers the relevant information from the whole collection and writes it into 16 tokens in the input space of a frozen LLM.
Results:
It beats Mem0, MemoRAG, and other baselines on five of eight benchmarks. On FactKG, it scores 84.33% against 80.43% with the full context.
It also works as agent memory. In TextWorld, using the agent’s own history as the collection raises success from 18.6% to 24.2%, and the agent revisits fewer rooms.
It also seems that it could help frozen LLMs accumulate and use knowledge effectively across long-context interactions. This is applicable to building continually learning AI systems.
On 18 October Elvis Saravia (@omarsar0) highlighted the paper “Memorilla: Latent Semantic Memory for LLMs” (arXiv 2610.21375, 15 October; Adibvafa Fallahpour, Jure Leskovec and co-authors). According to its abstract, the system writes information from a whole document collection into 16 query-specific memory tokens for a frozen LLM, beats baselines including Mem0 and MemoRAG on five of eight benchmarks, and on FactKG reaches 84.33% accuracy versus 80.43% with full-context inference while using 108x fewer tokens to represent the documents. The TextWorld success figures in the post (18.6% to 24.2%) are the poster’s summary. We include it as a research data point on the memory layer Chase calls super, super early.
Horizontal Agents Calling Vertical Ones
The second clip opens: “My take on the consumer space is that you will have both horizontal and vertical agents, and the horizontal agents will interact with the vertical agents.”
Chase’s reason is what each app already knows. “A lot of these apps have knowledge about you and can learn things about how you interact with that app better than a general horizontal agent could,” he said. He went on:
“It is a better experience if there is an agent that wraps those APIs and exposes a natural language face for a horizontal agent to interact with.”
Stebbings’ notes put the same view as: “Most apps should have a vertical agent because they can learn more about how a user interacts with them than a general agent can.”
Why the Harness Has to Stay Neutral
The notes link memory to a second argument, about the harness that connects a model to a company’s data and tools. “Companies want to be able to switch models, but the harness can be tightly tied to a model, and context today is intricately linked to the harness,” they say.
They continue: “If memory is built in a way that is linked to the harness, and the harness is linked to the model, the company ends up stuck.”
LangChain’s own About page says its commercial platform, LangSmith, gives companies control over “the layers that matter, like models, context, and deployment,” and that its newest product is an agent harness called Deep Agents.
What Agents Cost LangChain
The notes also give a figure for LangChain itself: it “spends roughly 30% of its engineering salary on coding tools, a share that has fallen over the last two months even as token usage has gone up.”
On large customers, the notes say: “Cost has risen as a concern for large enterprises, though it still ranks below accuracy and above latency.”
LangChain’s About page says it has crossed 1 billion open source downloads and ingests over 1 billion events per day on LangSmith.
The Structural Read
As we read it, the horizontal-and-vertical answer is a claim about who holds the user data: Chase says apps “can learn things about how you interact with that app better than a general horizontal agent could.”
As we read it, the harness passage in Stebbings’s notes is the commercial link: if memory is tied to the harness and the harness to the model, the notes say “the company ends up stuck.” LangChain’s About page says LangSmith gives companies control over models, context and deployment.
As we read it, the 30% figure in the notes shows the cost side from inside an agent company: it is a share of engineering salary that the notes say has fallen over two months while token usage went up.
Harrison Chase on 20VC, as posted by Harry Stebbings on X, 18 October 2026
“It is a better experience if there is an agent that wraps those APIs and exposes a natural language face for a horizontal agent to interact with.”
Three Implications
CONSUMER APP MAKERS Chase says “A lot of these apps have knowledge about you”, and Stebbings’s notes say most apps should have a vertical agent.
COMPANIES CHOOSING AN AGENT STACK Stebbings’s notes say companies want to be able to switch models, but “the harness can be tightly tied to a model.”
ENGINEERING LEADERS BUDGETING FOR AGENTS The notes say LangChain spends roughly 30% of its engineering salary on coding tools, with budgets the VP of engineering can raise.
The framework’s thesis reads: “The surface, not the model, governs autonomy. The agentic surface — the harness — is becoming where value pools in the entire AI stack.”
As we read it, Chase’s comments add memory to that surface: in Stebbings’s notes, memory built into the harness is what ties a company to a model. The framework was published in July 2026 and does not discuss LangChain or these comments.
What Is Not Established
From the two clips and the notes we read, it is not established how LangChain calculates the 30% figure. Nor is it clear which products beyond Instinct and Muse Chase counts as having a memory layer, or how horizontal and vertical agents would share what they learn about a user.
Business Engineer Framework
The Agentic Harness War
A Business Engineer framework on why the harness, not the model, is where value pools in the AI stack.
On 20VC, Chase said agent memory is “super, super early” and that horizontal agents will interact with vertical agents that wrap each app’s APIs. Stebbings’ notes tie memory to the harness and the harness to the model, a setup in which, they say, “the company ends up stuck.”
A note on sourcing. This piece is based on two clips of Harrison Chase that Harry Stebbings posted on X on 17 and 18 October 2026, Stebbings’s written notes from the conversation, which he says Chase allowed him to post, the YouTube description of the 20VC episode, and LangChain’s About page and Series B announcement. Quotes from the clips are as Stebbings posted them; we did not watch the full episode, and we did not contact Chase, LangChain or 20VC.
The 30% figure and the harness passage come from Stebbings’s notes. The references to our earlier reporting link to our own pieces.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
