Studies system kickoff

We introduced a new pattern in ai-labs: studies. A study is a curated reference collection — a directory of upstream specs, prior art, or reference implementations gathered before a decision is made or a feature is built. It is not a project (it ships nothing) and not documentation (it is living, checked-out code). The point of a study is to put real upstream implementations one cd away while we are deciding how to build something in the same domain.

The pattern is documented in two places. At the root of ai-labs the README now carries a short Studies section explaining what studies are and why we keep them. Inside ai-labs/studies/ there is a longer README.md covering the full concept (what a study is, why frontloading it improves decisions, why we reference it during development, and how to set one up properly), a CLAUDE.md that gives AI agents working in this repo norms for how to use studies (read from the pinned upstream rather than generating from training data; do not summarize study contents into prose unless asked; do not auto-bump pinned SHAs), and a context-v/prompts/Create-a-new-Study.md checklist that any future session can follow to spin a new study up from scratch in a repeatable way. The shape we settled on is: each study is its own GitHub repo under lossless-group/, named study-<topic-slug>, registered as a single submodule of ai-labs. Inside that repo, each upstream reference is a nested submodule — pinned to a specific SHA so what we read today is what someone reads in a month.

Two studies launched

Open Specs & Standards

A reading collection of open specifications, standards, and conventions that shape how humans and AI agents cooperate through files. The question we are investigating: when a person and an agent need to share context without ambiguity — which file conventions have already won the right to be the interface? Initial pins include AGENTS.md, Anthropic's Agent Skills (SKILL.md), the llms.txt proposal, Google Labs' DESIGN.md, the Model Context Protocol, the Agent2Agent (A2A) protocol, OpenSpec, GitHub's Spec Kit, the 12-Factor Agents principles, Frictionless Data, and a real MCP implementation (mcp-api from Parslee-ai) sitting alongside the spec. The intent is to extract patterns we can apply to our own agentic workflows — both developer and non-dev — without adopting any single one wholesale.

Memory Layers for Agents

A reading collection of production memory systems for AI agents — how they decide what to remember, when to update, how to scope, and what shape the memory artifacts take on disk. Initial pins are Mem0 (the most explicit "memory as a layer rather than a feature" implementation), and two contributions from Parslee-ai: neo (a self-improving code reasoning engine with persistent semantic memory) and statebench (a conformance test that measures state correctness over time). Pairing implementations with a benchmark in the same repo lets us evaluate the leading projects against a common harness instead of trusting each project's self-reported numbers. The README also captures a verified candidate list (Zep / Graphiti, Letta, LangMem, Memary, Cognee) ready to be added when the study expands.

Notes

Both study repositories are public on GitHub under lossless-group. External interest, suggestions, or contributions are welcome — reach out via the repos.