The models that week bragged about context windows. The number that actually bit was 20.

Nothing major is dated precisely to either day. Release trackers like llm-stats.com show the nearest dated open releases bracketing the window rather than landing on it: DiffusionGemma 26B-A4B from Google on June 10, and GLM-5.2 from Zhipu AI on June 16, billed as a top frontend-coding model. The reason a calendar-day search rarely returns a clean answer is that neither Hugging Face nor GitHub publishes a list keyed to a date. Both order discovery by a rolling trending score with no notion of when something shipped, so the date-honest feeds for any 48-hour window are huggingface.co/models sorted by trending, huggingface.co/papers, and github.com/trending.

Yes, broadly. The first half of June 2026 was one of the denser open-weight stretches on record, with a run of competitive Chinese releases (the Qwen, DeepSeek, MiniMax, Hunyuan, ERNIE, and GLM families all shipping new versions within a couple of weeks). MiniMax M3 landed June 1. GLM-5.2 landed June 16. But none of those carry a hard June 14 or June 15 timestamp, and a vendor's own benchmark numbers are self-reported until an independent harness re-runs them. Treat a fresh release as a thing to test, not a leaderboard fact.

Because context window is the spec that is easy to print and easy to compare: 256K, 512K, 1M tokens. It is a real capability, but it describes how much the model can read inside a single turn. It says nothing about whether your agent client will still have last Tuesday's conversation available to reopen. Those are two different limits. The model's context window is set by the vendor; the session-retention limit is set by the app you run the model inside, and it is the one that quietly decides whether your past work is reachable.

It is how many of your past conversations an agent client keeps in its history list and can restore. On June 15, 2026, Fazm raised that from 20 to 500. The change is a one-line edit to the SQL query in Desktop/Sources/MainWindow/Pages/ConversationHistorySection.swift, from LIMIT 20 to LIMIT 500, in commit 03499909. The inline comment records why: the 20 cap had been added on May 28 alongside a string-length performance fix, and it silently hid older chats from a power user (credited in the comment as Farid Mammadov) who had no search or pagination to reach them.

No, and the commit comment explains the reasoning rather than guessing. The performance storm that prompted the original cap was not row count; it was string LENGTH on large pasted prompts (40KB and up), and that was already fixed separately with SUBSTR(...,1,300) which only loads the first 300 characters of each preview. Row count is cheap because the history list is a SwiftUI LazyVStack that renders only the rows currently on screen. So 500 is a sane upper bound that stays until proper search and pagination land, without reintroducing the slowdown.

Auto-compacting happens inside a single live chat: the tool silently summarizes or drops earlier turns to keep the running context under a budget, which can erase decisions you made an hour ago. Retention is about chats you already closed: how many of them the app still lists and lets you reopen at all. Fazm addresses both. It does not auto-compact a live window, so the full history stays in context for the window's lifetime, and as of June 15 it lists up to 500 prior conversations instead of 20. A new model with a 1M-token window does not give you either property; the client does.

Open github.com/mediar-ai/fazm and read Desktop/Sources/MainWindow/Pages/ConversationHistorySection.swift. The query near the conversation summary load reads LIMIT 500, with a dated comment explaining the bump from 20. The git log shows commit 03499909, titled "Increase conversation history limit to 500," authored June 15, 2026. The changelog entry that followed (commit f9be2fd0) records the same fix. Everything is public, so you do not have to take any of this on faith.

Three feeds plus one habit. huggingface.co/models sorted by trending for weights and quantized variants. huggingface.co/papers for research that links an implementation. github.com/trending for the application layer of agents, MCP servers, and inference engines. The habit: keep the model you actually want to evaluate wired into the agent loop you use every day, with a session history deep enough that you can reopen the comparison you ran last week rather than starting over.