A model is a component, not a system.

The model will keep getting better on its own. The parts that decide whether you can trust it — memory, recovery, governance — are the parts we still build by hand.

_{— Vol. 01, §01 Thesis Field Notes}

This appendix records the source. Three systems, one concern each: persistence, recovery, orchestration. Everything below is either shipped, licensed, and installable — or explicitly marked as held.

№ 01  ·  memex — Persistence _{Python · MIT · ★12}

Persistent memory for AI coding agents over the Model Context Protocol: a bitemporal knowledge graph of your codebase, served to Claude Code, Cursor, Gemini CLI and any MCP client. Tree-sitter + Gemini Flash → Neo4j via Graphiti. Twelve MCP tools, hierarchical clusters, two-regime confidence decay.

10k+ interactions before recall degrades  ·  0 drift reconciled writes, no silent contradiction  ·  pip install memex-mcp

№ 02  ·  Cairn — Recovery _{Python · Apache-2.0 · ★2}

Recoverable long-horizon agents: a framework-agnostic reference harness plus a recovery-faithful live benchmark. The thesis — checkpoints are compactions — argues recovery is re-grounding, not replay.

100% resumable at last committed step  ·  <1s median time to re-attach and continue  ·  v1.0 held pending a powered live-LLM study

№ 03  ·  nexus — Orchestration _{Python · MIT · ★2}

A governed execution platform for autonomous agents: runtime orchestration, human approval workflows, sandboxed execution, recovery, and operational intelligence over the whole run.

4 layers agents · routing · policy · audit  ·  1:1 same input, same execution plan

_{FIG. 01 — How it fits.}

   request ──▶ nexus ──────────────────────────────▶ result
                 │  routing · policy · audit          ▲
                 │                                    │
                 ├──▶ memex   persistence  ───────────┤
                 │    bitemporal graph, Neo4j         │
                 │                                    │
                 └──▶ Cairn   recovery     ───────────┘
                      re-grounding, not replay

_{Smaller instruments, earlier volumes. Ordered by relevance, not stars.}

_{Also in circulation: antigravity · vision-to-action · gradia · MedMNIST-EdgeAI — the full index →}

Constraint first. If it cannot run on 6 GB of consumer VRAM, it is not finished — cloud GPUs are an optimisation, not a prerequisite.

Benchmarks before claims. Cairn is pinned at 0.x on purpose; the version number is a statement about evidence, not about effort.

Published means installable. PyPI, a license, and documentation, or it stays a field note.

Failure modes go in writing — quantisation deltas, recall decay curves, recovery latency. The number that embarrasses you is the one worth publishing.

Agent infrastructure   Model Context Protocol · bitemporal memory · knowledge graphs · durable execution · re-grounding recovery · policy & audit layers

Retrieval   RAG · Agentic RAG · LangGraph · vector search · Neo4j · Graphiti · Tree-sitter

Optimisation   GPTQ & 1–8 bit quantisation · distillation · pruning · ONNX Runtime · TensorRT · CUDA

Learning   PyTorch · TensorFlow · Transformers · CNNs · multimodal · OpenCV

Delivery   Python · TypeScript · FastAPI · Docker · GitHub Actions · Redis · GCP · AWS

Published   Transforming Urban Solutions for Smart Cities through Crowdsourced Feedback — March 2025

Certified   MCP Mastery, Fractal Analytics · RAG and Agentic AI, Coursera · Neural Networks with PyTorch, Coursera

_{FIG. 02 — Telemetry. 59 repositories, mostly Python. Since January 2024.}

editor:   Hill Patel — STiFLeR7
station:  CloudRedux · Pune, Maharashtra, India
volume:   01 — Persistent Systems
subject:  the architecture around the model
set in:   Python, mostly

hillpatel.tech  ·  LinkedIn  ·  X  ·  Medium  ·  Hugging Face  ·  stifler@hillpatel.tech