Every building gets a memory that heals itself.
Property management is context work: every email, invoice, bank transaction, owner question, contractor update, and meeting decision only makes sense when you know the building behind it. Today that knowledge is scattered across inboxes, PDFs, ERPs, bank exports, drives, and the memory of the property manager who has been there for years.
PropContext turns that chaos into one living building memory per property: a clear, source-backed knowledge file that people can inspect and AI agents can use immediately.
https://www.loom.com/share/39c4214527ed4dd5a4bfc712ab65483a
PropContext is a self-healing context layer for property management AI.
Instead of making an agent crawl every document again and again, PropContext compresses the important facts into a dense, readable property wiki. Each building gets an always-current memory of owners, tenants, contractors, open issues, invoices, decisions, obligations, and source references.
The inspiration is Karpathy's llm-wiki idea: a small, maintained markdown
knowledge base can be more useful and efficient than repeatedly searching a
large pile of raw documents. One compact building memory beats expensive RAG
gymnastics, chunk drift, and repeated context reconstruction.
PropContext does not regenerate the whole memory whenever something changes. It classifies new information, finds the exact place it belongs, patches only that section, preserves human notes, and keeps provenance back to the original source.
Hermes is the self-improvement loop behind the engine: it watches where ingestion misses facts, where schema gaps appear, and where agents need better context. From that feedback, Hermes can propose updates to extraction rules, wiki structure, vocabulary, and ingestion behavior so the system gets better dynamically instead of staying frozen after deployment.
AI agents for property management do not fail because they lack reasoning. They fail because they lack the right building context at the right time.
PropContext gives agents that context in the most efficient shape: one maintained, auditable memory per property. The human reads it, the agent acts from it, git remembers every change, and the ingestion engine keeps learning how to maintain it better.