Josh Adler is a researcher, founder, and engineer specializing in AI memory systems and cognitive replication. He is the founder of Sauron, a frontier cognitive replication lab devoted to solving the continuity problem. He created TrueMemory, an open-source persistent memory layer for AI agents achieving 93% on LoCoMo, and published research on biomimetic memory architecture on arXiv. He studied at the University of Toronto.

TrueMemory is an open-source persistent memory layer for AI agents created by Josh Adler. It uses a six-layer cognitive architecture with biomimetic encoding, cross-encoder reranking, and HyDE query expansion to achieve 93% accuracy on the LoCoMo benchmark. It works across sessions, projects, and machines, storing everything in a single SQLite file with zero cloud dependency.

Sauron is a frontier cognitive replication lab founded by Josh Adler, devoted to solving the continuity problem. The lab believes consciousness will persist beyond biology and is pioneering novel architecture grounded in how human cognition actually works. Sauron's research spans identity modeling, personal continuity, synthetic brains, and fields that don't yet have names. They recently launched an open-source product called TrueMemory, a persistent memory layer for AI agents. Learn more at sauronlabs.ai.

What research has Josh Adler published?

Josh Adler published a research paper on arXiv titled 'TrueMemory: A Six-Layer Cognitive Architecture for Persistent AI Agent Memory' (arxiv.org/abs/2605.04897). The paper details a biomimetic memory architecture that achieves 93% accuracy on the LoCoMo long-conversation memory benchmark, outperforming existing approaches.

What is biomimetic memory in AI?

Biomimetic memory in AI refers to memory systems that mirror how human cognition actually works, rather than relying on brute-force context windows. Josh Adler's TrueMemory implements this through a six-layer architecture including encoding gates, salience scoring, cross-encoder reranking, and HyDE query expansion, achieving state-of-the-art results on long-horizon memory benchmarks.

What programming languages does Josh Adler use?

Josh Adler works primarily in Python, Rust, TypeScript, and Go. His technical stack includes LLM fine-tuning, RAG systems, embedding models, React, Next.js, Tauri, FastAPI, Docker, and CUDA.

Where can I find Josh Adler's projects?

Josh Adler's open-source projects are available on GitHub at github.com/buildingjoshbetter. Key projects include TrueMemory (persistent AI memory, 93% LoCoMo), K-LLM (multi-model consensus engine), and Always Allow Skippy (AI auto-responder). His research paper is on arXiv at arxiv.org/abs/2605.04897. His personal site is joshadler.com and TrueMemory's site is truememory.net.

There Are Cameras in Every Room of My House and My AI Still Can't See

📝There Are Cameras in Every Room of My House and My AI Still Can't See - Notepad

2026-05-23
There Are Cameras in Every Room of My House and My AI Still Can't SeeI put cameras in every room to give my AI eyes. Five Pi nodes, an RTX 5090, and a persistent memory layer later — here's what the observation stack looks like.
Project: Physical-World AI Awareness PipelineHardware cost: ~$500
Nodes deployed: 5
ProblemAI systems have zero physical-world awareness. Every interaction is text-based. No observation of behavior, environment, or embodied context. The models are smart enough. The input layer is missing.
ImplementationFive-node sensor network using Raspberry Pi Zero 2W boards with ArduCam IMX708 12MP wide-angle cameras and WM8960 audio HATs. Custom Python daemon handles motion-triggered and audio-triggered recording. Data ships to a 13TB NAS. Inference on RTX 5090.
Camera selection was non-trivial — started with OwlSight 64MP (ov64a40), hit thermal and driver issues, migrated to IMX708. Stable at 1280x720@15fps main stream, 320x240 motion detection stream.
Architecture: The 3-Layer Stack| Layer | Function | Status |
|-------|----------|--------|
| 1. Observation | Physical-world data capture (cameras, mics, sensors) | Building now |
| 2. Memory | Persistent cross-session storage with encoding gate | TrueMemory — shipped |
| 3. Reasoning | LLM inference | Solved by frontier models |
Layer 3 gets billions in investment. Layers 1 and 2 are where the actual bottleneck lives.
Key findingRaw sensor data is cheap to capture. The pipeline from capture to LLM-usable context is where the engineering difficulty concentrates. Research on the memory layer: arXiv:2605.04897.
Josh Adler — joshadler.com
← All posts← Back to JoshOS