Local file-backed agent memory had a week, and the good ones publish verdicts.

Three projects inside seven days: daidocs with plain-text .dai files (claiming 92% on LongMemEval-S with Claude Fable 5, 22.40 points over the same model with no memory, at ~10x fewer tokens per question), Baron Munchausen at public alpha 0.6.1 with zero runtime dependencies, and Slowave on Show HN. What separates the first two from the pile is that both attach a judgment to the answer: Baron returns grounded, partial or ungrounded and names the unbacked sentences, and daidocs ships per-question judge verdicts with a sha256 manifest. Evaluating one of these, the retrieval claim matters less than whether it tells you when it's making things up.