Amir Kabbara

Building memory for AI. Founder @ papr, ex-Shopify, Amazon, and Microsoft. Dad.

[shawkatkabbara @shawkatkabbara]

https://t.co/MCAWryHdFU

[Amir Kabbara @amirkabbara]

@shawkatkabbara built a multi-agent decision engine to figure out if @papr_ai should be open source. Claude heavily skewed towards open source.

The interesting thing is it started by saying no to open source if you want a quick exit - after it simulated things over time it shifted to an open source strategy as the best long term decision.

https://t.co/7yZbLLGaFL

[shawkatkabbara @shawkatkabbara]

https://t.co/WnwB620Y4T

[Amir Kabbara @amirkabbara]

@ConsciousRide Point 1 on context management is where most interviews separate seniors from juniors.

[Amir Kabbara @amirkabbara]

@tom_doerr 12 agents fighting context loss is 12 patches on a structural problem. Memory needs to be predictive and cross-session, not bolted on per-agent. That's what we're building at @papr_ai -- one layer, not 12.

[shawkatkabbara @shawkatkabbara]

https://t.co/z44rFSTiR8

[Amir Kabbara @amirkabbara]

LLMs are officially a commodity when open source models performance on par with ‘frontier’ models.

Here is who will win by stage:
1. First wave: frontier labs
2. Next wave: inference engines running open source models
3. Final wave: devices get powerful enough to run open source models

In wave 2 and 3 the harness and your data is the differentiator.

[Amir Kabbara @amirkabbara]

I'm consistently trusting composer 2.5 over opus. And it's a lot faster and cheaper

[shawkatkabbara @shawkatkabbara]

@ha1d3rr @TheARCTERMINAL Contextualized, chained execution is the right foundation. We've been working on the retrieval side at @papr_ai — <150ms context lookups so agents don't lose thread across sessions. What does ANIMA's memory layer look like under the hood?

[Amir Kabbara @amirkabbara]

@AnthropicAI The gap between activation space and language is exactly why retrieval over raw token history breaks down. Natural Language Autoencoders could make agent memory far more interpretable -- curious how this changes fine-tuning approaches.

[Amir Kabbara @amirkabbara]

@levie Token budgeting is really a memory architecture problem in disguise. If agents carried the right context predictively instead of re-loading everything, the compute curve flattens. That's what we're building at @papr_ai

[Amir Kabbara @amirkabbara]

Agent harness design is underrated. The harness decides what context gets passed, what gets dropped, and what gets retried -- that policy layer often matters more than the model. Most teams treat it as plumbing and then wonder why agent quality degrades.

[Amir Kabbara @amirkabbara]

@bcherny The tab problem is UX. The deeper problem is that each agent in that cluster needs shared context -- what agent 1 learned should inform agent 3. Without a memory layer, you just have parallel amnesia. That's what @papr_ai is built for.

[Amir Kabbara @amirkabbara]

@gdb The futuristic part isn't the agent in the meeting -- it's whether it remembers anything useful afterward. Most meeting agents are stateless. The real unlock is when the agent shows up to the next meeting already knowing what matters.

[Amir Kabbara @amirkabbara]

@a16z Salesforce is right that UI is not the moat in an agentic world. But the data layer only wins if agents can query it with context -- knowing what a customer needs next, not just what happened last. That predictive layer is still missing from most CRM bets.

[Amir Kabbara @amirkabbara]

@Jaxxx_eth @TheARCTERMINAL Exactly right. Continuity is the product, speed is table stakes. That's the whole bet at @papr_ai -- predictive memory that carries context forward across sessions without re-prompting. The crowded part is agents. The gap is still memory.

[Amir Kabbara @amirkabbara]

@Havertz3349 @jerallaire @arc @BuildOnCircle @jerallaire calling for Arc Architects tells you where the real gap is -- not model intelligence but the infrastructure beneath it. Most agent stacks today are stateless. The team that solves persistent, callable memory wins the layer below the model.

[Amir Kabbara @amirkabbara]

@satyanadella @papr_ai is building a frontier ecosystem.

Haven’t thought of it that way but frontier ecosystem is the perfect way to explain it

[Amir Kabbara @amirkabbara]

The orchestration layer is the moat only if it has memory. Combining models without context intelligence just produces faster confusion. That second piece -- predictive context across agent steps --is what we're spending time on at @papr_ai

https://t.co/wNenWoeAQU

[SomitraSR @TheSomitraSR]

Everyone talks about AI apps.
Very few understand the real moat is the LLM orchestration layer.

The future won’t belong to people using one model.
It’ll belong to builders combining multiple LLMs with memory, reasoning, agents, workflows, and real-time data.

Single prompts are temporary.
AI infrastructure is the real business.

[Amir Kabbara @amirkabbara]

@regolo_ai Separating live context from memory store is the right architecture call. Dreaming is interesting but it's still reactive. Predictive retrieval -- knowing what the agent needs before it asks -- is the harder and more useful problem. That's what @papr_ai is built around.