Uncertainty is not just noise—it's the internal state that guides AI decision-making. In this episode of Memriq Inference Digest, we explore belief states, a foundational concept that enables AI systems to represent and reason about incomplete information effectively. From classical Bayesian filtering to cutting-edge neural planners like BetaZero, we unpack how belief states empower intelligent agents in real-world, uncertain environments.

In this episode:

- Understand the core concept of belief states and their role in AI under partial observability

- Compare symbolic, probabilistic, and neural belief state representations and their trade-offs

- Dive into practical implementations including Bayesian filtering, particle filters, and neural implicit beliefs

- Explore integrating belief states with CoALA memory systems for conversational AI

- Discuss real-world applications in robotics, autonomous vehicles, and dialogue systems

- Highlight open challenges and research frontiers including scalability, calibration, and multi-agent belief reasoning

Key tools/technologies mentioned:

- Partially Observable Markov Decision Processes (POMDPs)

- Bayesian filtering methods: Kalman filters, particle filters

- Neural networks: RNNs, Transformers

- Generative models: VAEs, GANs, diffusion models

- BetaZero and Monte Carlo tree search

- AGM belief revision framework

- I-POMDPs for multi-agent settings

- CoALA agentic memory architecture

Resources:

1. "Unlocking Data with Generative AI and RAG" by Keith Bourne - Search for 'Keith Bourne' on Amazon and grab the 2nd edition
2. This podcast is brought to you by Memriq.ai - AI consultancy and content studio building tools and resources for AI practitioners.

Discover how Recursive Language Models (RLMs) are fundamentally changing the way AI systems handle ultra-long contexts and complex reasoning. In this episode, we unpack why RLMs enable models to programmatically query massive corpora—two orders of magnitude larger than traditional transformers—delivering higher accuracy and cost efficiency for agentic AI applications.

In this episode:

- Explore the core architectural shift behind RLMs and how they externalize context via sandboxed Python environments

- Compare RLMs against other long-context approaches like Gemini 1.5 Pro, Longformer, BigBird, and RAG

- Dive into technical trade-offs including latency, cost variability, and verification overhead

- Hear real-world use cases in legal discovery, codebase analysis, and research synthesis

- Get practical tips on tooling with RLM official repo, Modal and Prime sandboxes, and hybrid workflows

- Discuss open challenges and future research directions for optimizing RLM deployments

Key tools and technologies mentioned:

- Recursive Language Model (RLM) official GitHub repo

- Modal and Prime sandboxed execution environments

- GPT-5 and GPT-5-mini models

- Gemini 1.5 Pro, Longformer, BigBird architectures

- Retrieval-Augmented Generation (RAG)

- Prime Intellect context folding

- MemGPT, LLMLingua token compression

Timestamps:

00:00 - Introduction to Recursive Language Models and agentic AI

03:15 - The paradigm shift: externalizing context and recursive querying

07:30 - Benchmarks and performance comparisons with other long-context models

11:00 - Under the hood: how RLMs orchestrate recursive sub-LLM calls

14:20 - Real-world applications: legal, code, and research use cases

16:45 - Technical trade-offs: latency, cost, and verification

18:30 - Toolbox and best practices for engineers

20:15 - Future directions and closing thoughts

Resources:

"Unlocking Data with Generative AI and RAG" by Keith Bourne - Search for 'Keith Bourne' on Amazon and grab the 2nd edition

This podcast is brought to you by Memriq.ai - AI consultancy and content studio building tools and resources for AI practitioners.

Stay tuned and keep pushing the boundaries of AI engineering with Memriq Inference Digest!

# Evaluating Agentic AI: DeepEval, RAGAS & TruLens Frameworks Compared

In this episode of Memriq Inference Digest - Engineering Edition, we explore the cutting-edge evaluation frameworks designed for agentic AI systems. Dive into the strengths and trade-offs of DeepEval, RAGAS, and TruLens as we unpack how they address multi-step agent evaluation challenges, production readiness, and integration with popular AI toolkits.

In this episode:

- Compare DeepEval’s extensive agent-specific metrics and pytest-native integration for development testing

- Understand RAGAS’s knowledge graph-powered synthetic test generation that slashes test creation time by 90%

- Discover TruLens’s production-grade observability with hallucination detection via the RAG Triad framework

- Discuss hybrid evaluation strategies combining these frameworks across the AI lifecycle

- Learn about real-world deployments in fintech, e-commerce, and enterprise conversational AI

- Hear expert insights from Keith Bourne on calibration and industry trends

Key tools & technologies mentioned:

DeepEval, RAGAS, TruLens, LangChain, LlamaIndex, LangGraph, OpenTelemetry, Snowflake, Datadog, Cortex AI, DeepTeam

Timestamps:

00:00 - Introduction to agentic AI evaluation frameworks

03:00 - Key metrics and evaluation challenges

06:30 - Framework architectures and integration

10:00 - Head-to-head comparison and use cases

14:00 - Deep technical overview of each framework

17:30 - Real-world deployments and best practices

19:30 - Open problems and future directions

Resources:

1. "Unlocking Data with Generative AI and RAG" by Keith Bourne - Search for 'Keith Bourne' on Amazon and grab the 2nd edition
2. This podcast is brought to you by Memriq.ai - AI consultancy and content studio building tools and resources for AI practitioners.