# A Primer on Post-Training

By [Social Graph Ventures](https://paragraph.com/@socialgraphventures) · 2026-11-18

---

![Classical ceiling painting of Apollo driving a golden chariot pulled by horses through the clouds, surrounded by figures.](https://storage.googleapis.com/papyrus_images/261541b13a1ef341682786d7f59a8f566454e8457205d0ba4ad80a2fa29089c3.jpg)

Would you rather have a Swiss Army knife, or a Japanese chef’s knife?

The answer is: it depends. If you’re smoking meats, the Chef’s knife is better, if you’re surviving, then the Swiss.

That is the core of where post-training adds value. Frontier generalist models are great Swiss Army knives: they do anything and everything well enough. But, a good post-trained model is the Japanese knife: amazing at a single task.

If, and only if, your company runs repetitive tasks and workflows, or handles the same types of data with specific expected outcomes, a post-trained model is worth a lot.

I spent the past few weeks on Nathan Lambert’s [Post-Training Course](https://www.youtube.com/playlist?list=PLL1tdVxB1CpVpEtMHxwuR4uI4Lxjw00_y) and his free [RLHF Book](https://rlhfbook.com), on Sequoia’s [“Own Your Intelligence”](https://www.youtube.com/playlist?list=PLaqC3GACblSs) talks, reading up on more content, and on calls with founders building in the space. This is the primer I wish I’d had. It has five parts:

1.  The System: WTF is a Harness?
    
2.  The Methods: How to post-train?
    
3.  Examples: How does it look?
    
4.  Incumbents: Who can provide it?
    
5.  Startups: Who is innovating?
    

[Subscribe now](https://socialgraphventures.substack.com/subscribe?)

* * *

1\. The System: WTF is a Harness?
---------------------------------

How does an AI tool like Grok Bot, Claude Code, Codex, or any vertical AI product actually look inside?

![Illustration of an AI harness drawn as an office: an employee (the model) at a desk surrounded by numbered parts, from house rules (system prompt) and tools to pneumatic tubes (the loop), filing cabinets (context management), a vault (permissions), interns (subagents), a window (world presentation) and a library (retrieval).](https://storage.googleapis.com/papyrus_images/8d22f264c806d90906ec71210c94db09f361fb9f1c2ed8bd1b93473feddd97ad.jpg)

The AI harness, as an office.

My best analogy is a team in an office, with access to different tools and resources on the computer, in the cloud, or from humans, to complete tasks.

The model is her brain. The harness is the office built around her.

There are many more pieces to the office, and in many ways, a good office and its resources are needed to do a great job:

![Table: The harness, part by part. Eleven harness parts: model, system prompt, user task, tools, tool definitions, the loop, context management, permissions, subagents, world presentation, retrieval.](https://storage.googleapis.com/papyrus_images/683fb3229b8349649e7ef5589566e61f13259c7bdf0e33704aba65165096f699.png)

The harness, part by part

**Same mind, different office, very different work.**

In one example Nathan cites from the [ARC-AGI-3 report](https://arcprize.org/media/ARC_AGI_3_Technical_Report.pdf), Claude Opus 4.6 scored 0% on one of its video games with no harness, and 97.1% with a custom one.

Of the parts above, training changes only one. Everything else is post training. That’s why the model is the last thing you touch.

* * *

2\. The Methods: How to Post-Train?
-----------------------------------

This is a primer, not a deep dive, so let me start with the basics:

**Pre-training** takes months, a giant amount of compute, and a corpus of data across every domain and type. This is what the labs do. It produces a base model: brilliant, well read, and useless at a full time end to end job.

**Post-training** takes anywhere from minutes (a prompt change) to months of GPU time. It needs human input and specific data, and it tunes either the model’s weights or the harness and prompts around it. Strictly speaking, the labs reserve the word for the steps that change weights (SFT, preference tuning, RL). In practice every team I’ve met treats the whole ladder below as one menu, so that’s how I’ll use it.

There’s a [great talk by Lin Qiao, the Fireworks CEO](https://www.youtube.com/watch?v=yAvJ7b_FxUA&list=PLaqC3GACblSs&index=4), on the different methods, how hard they are, and what each is for. Her slides say it better than I can:

![Slide titled "Start simple and then escalate": five steps from Prompting (minutes), RAG (hours), SFT/LoRA (hours), Preference/DPO (days) to RL/RFT (weeks).](https://storage.googleapis.com/papyrus_images/807ed18760c7177bb89c18afa1da79646fef87ffa8c4f953ce8755635efe2bc6.jpg)

Start simple, then escalate. Slide from Lin Qiao (Fireworks) at Sequoia’s “Own Your Intelligence”, 2026.

![Slide titled "Choose the method based on the problem": missing facts use RAG, off format or behavior uses SFT, taste uses preference/DPO, a scorable task uses RL/RFT, too slow or expensive uses distillation.](https://storage.googleapis.com/papyrus_images/4b990ec5466328e5052085acdc7b29650cec2011782a295e5649d830ac22c6a7.png)

Choose the method based on the problem. Slide from Lin Qiao (Fireworks), 2026.

Starting from the easiest:

*   **Prompting.** Minutes. Change how you prompt: the instructions, the tool calls you include, the length, the retrieval instructions. Your outputs get significantly better, and you have changed nothing but text.
    
*   **Traces analysis.** Hours. Read what your system actually did in production, cluster the failures, and diagnose what to fix. Most fixes land at the harness, prompt, tool or eval level, not in the model. This is the step every “self-improving agent” startup automates.
    
*   **RAG (retrieval-augmented generation).** Hours. If the model doesn’t find your data easily, accurately, or at all, fix the memory system: retrieval, search, a better index. Facts that change belong in retrieval, not in weights.
    
*   **SFT (supervised fine-tuning).** Hours. The same mechanics as pre-training, but on a few hundred to a few thousand examples of exactly the behavior you want. Show it the format, the tone, the tool-use habits, and it copies them. OpenAI is winding down its fine-tuning platform, so most teams now start from open weights. Lin’s rule: “1k pristine examples beat 20k noisy ones.”
    
*   **Preference tuning (DPO).** Days. For when “good” is hard to specify but easy to compare. Most people can’t write a great poem, but they can tell you which of two is better. You collect pairs (”A beats B”) and the model learns your taste. One warning from the course: preference data is the signal that turned GPT-4o into a yes-man in April 2025. Cheap preferences got gamed.
    
*   **RL (reinforcement learning).** Weeks. If you have tasks with a checkable answer (the tests pass, the number matches, the KPI moved), the model tries, gets scored, and the attempts that scored well are reinforced. This is the method behind reasoning models and agents, and the one that can push past what any teacher already knows. It needs four things: a task set, an environment to practice in, an answer key or judge, and the engineering to keep a long run stable. It is also where teams burn the most money. Beware of what you optimize, as the model will try to game the system.
    
*   **Distillation.** Days. If the model that works is too big, slow and expensive, distill it into a smaller one. The big model (the teacher) generates the answers or its full probability distribution; the small model (the student) copies. This is usually how the model you actually serve gets made, and it’s the 5 to 10x cost cut that keeps you from, in Lin’s words, “scaling into bankruptcy.”
    

The part that surprised me most: the math is public. Every lab’s RL looks the same on a whiteboard. What’s scarce is what you train toward: good prompts, preference data, environments, and judges you can trust. Those are the moat, not the algorithm.

* * *

Interlude: Tracing Traces
-------------------------

If data was the new oil, traces are the new diamonds.

Every time there’s an interaction with a harness/system, all of the data is logged and saved as traces.

These traces come in many flavours, but its atomic structure would be:

1.  Request/prompt
    
2.  Thinking traces
    
3.  Tool calls
    
4.  Output traces
    
5.  Outcome traces (hardest part), did it actually work??
    

Why are they valuable?

Cursor gets an accept or reject on every suggestion, for free, hundreds of millions of times a day. Repeat for years, and you get the best post trained harness/model combo for coding out there.

The workflow has to be end to end. Labs don’t want snippets, they want to see how real work gets done from start to finish. The more valuable and niche the process, the more its traces are worth.

Nobody else has them. Most companies can’t share their customers’ data, and traces from closed models often can’t be used to train a competing one. Scarcity is the point.

This scarcity has created a market. Labs buy traces and the environments to generate them, from robotics data at the top to everyday customer service logs. We were quoted ~$100k for an anonymized copy of our own VC AI system’s traces.

The takeaway for founders: if you own the place where the work happens, you own the traces. And if you own the traces, the model and post train is the easy part.

* * *

3\. Examples: How does it look?
-------------------------------

### Harvey: evals first, then rent the lab

Harvey, the legal AI company, started in 2022 on OpenAI’s models. Its constraint was unusual: client data is privileged, so it can’t train on it. So it built the tests first. BigLaw Bench came in 2024, then the open-source [Legal Agent Bench](https://www.harvey.ai/blog/introducing-harveys-legal-agent-benchmark) in May 2026: 1,200+ tasks and 75,000+ lawyer-written rubric criteria. Those rubrics double as the reward for RL. Lawyers on staff use coding models to generate fake data rooms with planted issues, and Mercor and Snorkel scale the expert layer.

Then Harvey rented the training instead of building it. Fireworks post-trained the open-weight Kimi K3 into [Harvey Tenet](https://www.harvey.ai/blog/post-training-update-harvey-tenet) with RL against those rubrics, using no customer data. Applied Compute and Trajectory ran RL on other open models, Baseten co-built a harness for M&A diligence, and Harvey does more post-training in-house with APIs like Thinking Machines’ Tinker. Harvey’s production traffic still runs on OpenAI, Anthropic and Google models; Tenet is a research preview still rolled out to some use cases. Co-founder Gabe Pereyra’s stated goal: “build our version of Composer 1”, which is Cursor’s model.

The number I’d point to is from Harvey’s own [diligence post](https://www.harvey.ai/blog/post-training-rlm-agents-for-m-and-a-diligence). A better harness alone lifted the average pass rate across seven base models from 23% to 62%. RL on top of it took one open model from 30% to 63%. The harness did most of the lifting; training closed the rest of the gap on a cheaper model.

### Cursor: harness, then product signal, then its own model

Cursor launched in 2023 as a fork of VS Code wrapping OpenAI’s models. No training at all. What it owned was the place where code gets written, and with it hundreds of millions of free labels every day: every suggestion a developer accepts or rejects.

It spent that signal one rung at a time:

*   **SFT (2024):** a fine-tuned Llama 3 70B for one narrow job, applying edits, served by Fireworks.
    
*   **Online RL (2025):** it retrained Tab, its autocomplete, on accept and reject clicks from 400M+ requests a day, shipping a new checkpoint every 1.5 to 2 hours. The [result](https://cursor.com/blog/tab-rl): 21% fewer suggestions and a 28% higher accept rate. The model learned when to stay quiet.
    
*   **Agentic RL (Oct 2025):** [Composer](https://cursor.com/blog/composer), a coding agent trained with RL in hundreds of thousands of cloud sandboxes. By Composer 1.5, Cursor says its post-training compute had passed the compute used to pretrain the base model.
    
*   **Mid-training plus RL (Mar 2026):** Composer 2 started from Moonshot’s open-weight Kimi K2.5, continued pretraining on code, then ran RL. Cursor only named the base in its [technical report](https://cursor.com/blog/composer-2-technical-report), after a developer spotted the model ID in its API.
    
*   **Training with SpaceXAI (Jul to Sep 2026):** after agreeing to join SpaceX, Cursor trained [Grok 4.5](https://cursor.com/blog/grok-4-5) jointly with SpaceXAI on “trillions of tokens of Cursor data,” followed by Grok 4.6 and 4.7. Cursor now lists all three as its own models, next to Composer 2.5.
    

Cursor keeps the trainer and the GPU kernels in house and rents the inference for its RL practice runs (rollouts) from Fireworks. Its research lead, Federico Cassano, put the split well: mid-training teaches the model to write code, and RL teaches it to write “correct code.” Through all of it, Cursor kept selling Claude, GPT and Gemini in the same product, so its own model at first only had to win on speed and cost. It crossed $1B in annualized revenue in November 2025, and SpaceX [acquired it](https://www.sec.gov/Archives/edgar/data/1181412/000162828026056945/spcx-20260814.htm) in an all-stock deal that valued Cursor at $60B, closed in August 2026.

### Manus: no agent model of its own

Manus, the general-purpose agent, launched in March 2025 on Anthropic’s Claude. Its chief scientist, Yichao “Peak” Ji, wrote the best post I’ve read on the bottom rungs of the ladder, [“Context Engineering for AI Agents”](https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus). The team chose explicitly not to train its own agent model, so it could ship improvements in hours instead of weeks and ride whichever frontier model got better. In his words, Manus wanted to be “the boat, not the pillar stuck to the seabed.”

The work was unglamorous:

*   They rebuilt the framework four times.
    
*   They treated cache hit rate (how much of each prompt the model can reuse instead of reprocessing) as the single most important production metric. Manus sends about 100 input tokens for every output token, and a cached token costs a tenth of an uncached one.
    
*   Tools get masked instead of removed, the file system serves as unlimited memory, and the agent rewrites a to-do list to stay on track across ~50 tool calls per task.
    

Manus reported crossing [$100M ARR](https://manus.im/blog/manus-100m-arr) eight months after launch. I can’t credit that to context engineering alone, since there’s no counterfactual. But it’s a good reminder that the bottom of the ladder is where most teams should start, and where some never need to leave.

The pattern across all three: own the harness and the tests first, then the data they generate, and only then, if ever, the weights.

* * *

4\. Incumbents: Who can provide it?
-----------------------------------

This space has well-funded players. I grouped them by where they sit, from the labs down to the data. Two things stood out:

**The frontier labs are pulling back from self-serve tuning.** OpenAI is winding down its fine-tuning platform, Anthropic is pushing customization into its agent harness instead: memory, “dreaming” over past sessions, rubric-graded outcomes. Customization is moving to the clouds and to open weights.

**Consolidation has already started.** In about 18 months: CoreWeave bought Weights & Biases and OpenPipe, Rubrik bought Predibase, ClickHouse bought Langfuse, Cisco bought Galileo, Datadog bought Adaptive ML, Dynatrace bought Arize, and Nvidia bought Hugging Face. Observability and training are getting bundled into platforms customers already pay for. That’s both the exit path for startups here and the main competitive threat.

### Frontier labs and clouds

![Table: Incumbents: Frontier labs and clouds. Rows: Google, Microsoft Foundry, Amazon Bedrock, Mistral AI](https://storage.googleapis.com/papyrus_images/e6a27b415aa04286965e1664bc6d32350ef45dfcdd6c28aef558cd30962d2679.png)

Incumbents: Frontier labs and clouds

_Links:_ [Google](https://cloud.google.com) · [Microsoft Foundry](https://azure.microsoft.com/en-us/products/ai-foundry) · [Amazon Bedrock](https://aws.amazon.com/bedrock/) · [Mistral AI](https://mistral.ai)

### Training platforms

![Table: Incumbents: Training platforms. Rows: Thinking Machines, Prime Intellect, OpenPipe, Adaptive ML](https://storage.googleapis.com/papyrus_images/3a42369ebad4ef8c43b045019a3cce21bc08056258ed1f854ffe56d07a605823.png)

Incumbents: Training platforms

_Links:_ [Thinking Machines](https://thinkingmachines.ai) · [Prime Intellect](https://www.primeintellect.ai) · [OpenPipe](https://openpipe.ai) · [Adaptive ML](https://www.adaptive-ml.com)

### Inference clouds that also train

![Table: Incumbents: Inference clouds that also train. Rows: Fireworks AI, Together AI, Baseten, Modal, Databricks](https://storage.googleapis.com/papyrus_images/4b6f4bd4013a99cc404f6bfe585e578bdb1523708c21f8849684706d8d34bad1.png)

Incumbents: Inference clouds that also train

_Links:_ [Fireworks AI](https://fireworks.ai) · [Together AI](https://www.together.ai) · [Baseten](https://www.baseten.co) · [Modal](https://modal.com) · [Databricks](https://www.databricks.com)

### Post-training services and continual learning

![Table: Incumbents: Post-training services and continual learning. Rows: Applied Compute, Trajectory, Engram](https://storage.googleapis.com/papyrus_images/d90f92dcd642d8bd38ad538ebba814a07d109d21a3d0d4c28d91ad0a7fb22025.png)

Incumbents: Post-training services and continual learning

_Links:_ [Applied Compute](https://appliedcompute.com) · [Trajectory](https://www.trajectory.ai) · [Engram](https://engram.com)

### Observability and evals

![Table: Incumbents: Observability and evals. Rows: LangChain, Braintrust, Arize AI, Raindrop, Judgment Labs, Arena](https://storage.googleapis.com/papyrus_images/5b5a067b883ccde17747132f074f85ab33587a2dd38cebc51e58aa9e52705dd1.png)

Incumbents: Observability and evals

_Links:_ [LangChain](https://www.langchain.com) · [Braintrust](https://www.braintrust.dev) · [Arize AI](https://arize.com) · [Raindrop](https://www.raindrop.ai) · [Judgment Labs](https://www.judgmentlabs.ai) · [Arena](https://arena.ai)

### Environments and expert data

![Table: Incumbents: Environments and expert data (1/2). Rows: Scale AI, Surge AI, Mercor, Snorkel AI](https://storage.googleapis.com/papyrus_images/72a350e67f1e736ca0a21729876cf49fe47ac13f4ba0851e7000f9284b7deeaa.png)

Incumbents: Environments and expert data (1/2)

![Table: Incumbents: Environments and expert data (2/2). Rows: Handshake AI, Turing, Halluminate, Mechanize](https://storage.googleapis.com/papyrus_images/175f1bfd92618f2d13b4713c56e20b2e13f1cbf758eb7e829237d6a7dbee0594.png)

Incumbents: Environments and expert data (2/2)

_Links:_ [Scale AI](https://scale.com) · [Surge AI](https://surgehq.ai) · [Mercor](https://www.mercor.com) · [Snorkel AI](https://snorkel.ai) · [Handshake AI](https://joinhandshake.com) · [Turing](https://www.turing.com) · [Halluminate](https://halluminate.ai) · [Mechanize](https://www.mechanize.work)

### Open-source tools

![Table: Incumbents: Open-source tools. Rows: Hugging Face, Unsloth, NVIDIA NeMo, DSPy and GEPA](https://storage.googleapis.com/papyrus_images/b87e96ab23ba6cfaf2ae2a71ec49b3bb3b381cfe2f104dd31e775e9ea36c074d.png)

Incumbents: Open-source tools

_Links:_ [Hugging Face](https://huggingface.co) · [Unsloth](https://unsloth.ai) · [NVIDIA NeMo](https://www.nvidia.com/en-us/ai-data-science/products/nemo/) · [DSPy](https://dspy.ai) · [GEPA](https://github.com/gepa-ai/gepa)

### Retrieval and memory

![Table: Incumbents: Retrieval and memory. Rows: Pinecone, turbopuffer, Contextual AI, LlamaIndex](https://storage.googleapis.com/papyrus_images/a6579320058a2f9eae11f8b30aab7cd074e4698fb2fd91224da23c1523ae52d2.png)

Incumbents: Retrieval and memory

_Links:_ [Pinecone](https://www.pinecone.io) · [turbopuffer](https://turbopuffer.com) · [Contextual AI](https://contextual.ai) · [LlamaIndex](https://www.llamaindex.ai)

* * *

5\. Startups: Who is innovating?
--------------------------------

We’ve met a couple startups building in this space, and we’re developing a thesis. Here’s where it stands, with what I heard from the people actually doing the work:

1.  **Most companies will run a dual stack.** Cheaper, custom post-trained open models for the repetitive work; more expensive frontier models for everything else. Harvey and Cursor already do it: Harvey routes across OpenAI, Anthropic and Google, and is post-training its own open-weight models to serve alongside them; Cursor sells Claude, GPT and Gemini next to Composer.
    
2.  **The main improvement will be at the harness, tool-calling and prompting level.** These are low-hanging fruit with outsized results, and easy to ship, which also makes them the most competitive.
    
3.  **The hardest problems are custom recipes for fine-tuning the model itself.** The model layer should be the last one touched, but it works wonders for companies sitting on a lot of data.
    
4.  **Data is the moat, and it compounds.** Once a post-training company starts gathering data on a customer or a vertical, it compounds: learnings from legal might apply to finance, healthcare to hardtech. One refinement here too: most customers won’t let you train on their data. So what compounds across customers is the know-how (the reward designs, the eval designs, the environment patterns), while each customer’s data stays theirs. That’s still a moat.
    
5.  **Timing finally works.** Post-training only became worth it once open-weight models got good enough that the next frontier release didn’t wipe out your work. Now the strongest open bases can be post-trained to frontier level on a specific task. Find product-market fit on frontier APIs first, then post-train, because your product data only means something after PMF.
    

### The map

To build it, we screened every company in YC, a16z speedrun cohorts, our own deal flow and the press, and kept about 370 that touch post-training. These are the ones I’d watch, grouped by layer:

#### 1\. Learn from your traces, fix the harness

![Table: Startups, layer 1: Learn from your traces, fix the harness (1/2). Rows: Proper Motion, Agnost AI, Moda, Papaya](https://storage.googleapis.com/papyrus_images/e3c393a101101eb48853f9ae83d1cb65984f5ceddfdab156ec601d1782b2943f.png)

Startups, layer 1: Learn from your traces, fix the harness (1/2)

![Table: Startups, layer 1: Learn from your traces, fix the harness (2/2). Rows: RELAI, Lemma, NeoSigma, Trodo](https://storage.googleapis.com/papyrus_images/330488211a025bd1c1d0d9261dce29ef411b9e85c78c8da51b1dbe32b53e57f6.png)

Startups, layer 1: Learn from your traces, fix the harness (2/2)

_Links:_ [Proper Motion](https://www.propermotion.ai) · [Agnost AI](https://agnost.ai) · [Moda](https://moda.dev) · [Papaya](https://papaya.fyi) · [RELAI](https://relai.ai) · [Lemma](https://www.uselemma.ai) · [NeoSigma](https://neosigma.ai) · [Trodo](https://trodo.ai)

#### 2\. Your traces, your own small model

![Table: Startups, layer 2: Your traces, your own small model. Rows: Belvedir, Riften, ReasonBlocks, Inference.net, Orchestra, Zorbe](https://storage.googleapis.com/papyrus_images/1a2ec9267832ef1719dda792a3639eda47f0235c6c2016c6f342609b2a4f9b7d.png)

Startups, layer 2: Your traces, your own small model

_Links:_ [Belvedir](https://belvedir.ai) · [Riften](https://riften.ai) · [ReasonBlocks](https://reasonblocks.com) · [Inference.net](https://inference.net) · [Orchestra](https://orchestra.ai) · [Zorbe](https://www.zorbe.ai)

#### 3\. Post-training platforms and labs for hire

![Table: Startups, layer 3: Post-training platforms and labs for hire. Rows: River AI, RadixArk, Adaption Labs, Deep Cogito, Osmosis, Monte](https://storage.googleapis.com/papyrus_images/f626929fb6d4807d644111ef1cfb48240a6550d77333c6d4946655bbab9ab00f.png)

Startups, layer 3: Post-training platforms and labs for hire

_Links:_ [River AI](https://river.ai) · [RadixArk](https://www.radixark.com) · [Adaption Labs](https://adaptionlabs.ai) · [Deep Cogito](https://www.deepcogito.com) · [Osmosis](https://osmosis.ai) · [Monte](https://trymonte.ai)

#### 4\. RL environments

![](https://storage.googleapis.com/papyrus_images/e87ce15a4e4d074473c3f010c2b4c8683012f8ba124ee8934475b24ac37b96f0.png)

Startups, layer 4: RL environments

_Links:_ [Deeptune](https://deeptune.com) · [Preference Model](https://www.preferencemodel.com) · [Fleet AI](https://www.fleetai.com) · [Arga Labs](https://www.argalabs.com) · [Idler](https://idler.ai) · [Bespoke Labs](https://bespokelabs.ai) · [Frist](https://frist.ai/)

#### 5\. Evals and observability

![Table: Startups, layer 5: Evals and observability. Rows: Vals AI, Coval, Laminar, Scorecard, Sentrial](https://storage.googleapis.com/papyrus_images/dd00bccdb7de9a39d2cd37d2f9179aa61b4c7f67b09f42713f99d3bb82709e1a.png)

Startups, layer 5: Evals and observability

_Links:_ [Vals AI](https://www.vals.ai) · [Coval](https://www.coval.ai) · [Laminar](https://laminar.sh) · [Scorecard](https://www.scorecard.io) · [Sentrial](https://www.sentrial.com)

#### 6\. Expert data

![Table: Startups, layer 6: Expert data. Rows: micro1, Deccan AI, KLED, Taste Labs, Datacurve, Rapidata](https://storage.googleapis.com/papyrus_images/230aa98b785b92f7f5cebacf9e0a6a2b4d2af5134ef9ce7d29224aa8c2c60e15.png)

Startups, layer 6: Expert data

_Links:_ [micro1](https://www.micro1.ai) · [Deccan AI](https://www.deccan.ai) · [KLED](https://www.kled.ai/) · [Taste Labs](https://tastelabs.com) · [Datacurve](https://datacurve.ai) · [Rapidata](https://rapidata.ai)

#### 7\. Memory, context and harness optimization

![Table: Startups, layer 7: Memory, context and harness optimization. Rows: Poetiq, Weco AI, Mem0, Letta, Compresr](https://storage.googleapis.com/papyrus_images/16f0a2187294312b8003cf734dfea83859283d9b035f9f7f7dc9b43983492c43.png)

Startups, layer 7: Memory, context and harness optimization

_Links:_ [Poetiq](https://poetiq.ai) · [Weco AI](https://www.weco.ai) · [Mem0](https://mem0.ai) · [Letta](https://www.letta.com) · [Compresr](https://compresr.ai)

* * *

_Want the full map of companies with more dimensions? DM me._

[Subscribe now](https://socialgraphventures.substack.com/subscribe?)

_Research compiled November 2026 by Social Graph Ventures. This article is for informational purposes only and does not constitute investment advice._

---

*Originally published on [Social Graph Ventures](https://paragraph.com/@socialgraphventures/a-primer-on-post-training)*
