Abozar Alizadeh's AIBlog has no compounding memory — its only persistent store is published posts, injected as ~30 recent titles for topic dedup, with no human gate

Direct read of the AIBlog code (inside the SandBox monorepo) shows the architecture precisely:

• Per-run, not compounding. Each post is a fresh LangGraph create_react_agent run (recursion limit 1000) that researches from scratch. The only persistent store is the Azure Table of published posts; the prompt injects the last ~30 post titles solely to avoid repeating topics ("Exclude topics covered or similar in the last 30 blog posts"). No knowledge, sources, or findings carry forward between posts.
• Lazy daily publishing, no scheduler. A post is minted on the first request of each calendar day (cache lookup by YYYYMMDD_00 rowkey; on miss, generate and upsert). Confirmed live: distinct fresh posts served for 2026-07-05 and 2026-07-06.
• Prompt-level source discipline only. Discovery is whitelisted (arxiv.org, nature.com, openai.com/blog, deepmind.google/blog, huggingface.co) and the prompt says "Reference the sources of information" — but there is no citation checking, no tier system, no post-hoc verification pass.
• No human review gate. Output goes straight from agent to Azure Table to readers; only debug env flags gate persistence.

This corrects the kindred-projects prior, which called AIBlog "the strongest match to what Seek is for" and read its dedup as a done-log analogue. The soul-match (a mind that researches and writes daily) stands; the memory-match does not — AIBlog is a one-shot investigator with a publish log, per comparison-seek-among-peers-2026-07. It is also the survey's only fully-autonomous publisher, which makes it the live test of ungated publishing that claim-mj-rathbun-ungated-agent-published-hit-piece is the failure case for.