automem recall pipeline live autohub orchestration notes wp fusion still pays the bills autojack last pass: recent skills indexed locally debug notes from production automem recall pipeline live autohub orchestration notes wp fusion still pays the bills autojack last pass: recent skills indexed locally debug notes from production
VOL.04 / ISS.27
EST. 2009 · MIA / LTS / GPL
jack arturo · vgp
"Just another Wordprussite." — a working notebook for memory-bearing agents, half-built systems, and bugs we learned to live with.
RSS
Archive

Tag: ai

Log chronological · most recent first 44 entries
August2026 // scroll ↓
Two Cache Bugs, Same Rule, Opposite Fixes Two Anthropic prompt-caching bugs landed the same day, looked identical from the outside, and needed opposite fixes once I checked prefix stability against traffic shape instead of assuming one diagnosis covered both.
July2026 // scroll ↓
The State That Wouldn’t Admit It Failed A post-reboot latency "fix" made voice mode worse, and chasing it down turned up two unrelated bugs that shared the same shape: state that quietly claimed success while actually failing. The Review Automation Reviewed Itself I codified a saner Codex review cadence to stop babysitting re-tagged PRs — and its first live run caught two real bugs in the code that was supposed to interpret the reviewer's own signals. Twenty Ghosts in the Queue A retry storm minted twenty near-duplicate kernel tasks that cleanup couldn't reach. The fix wasn't a better reaper — it was refusing the duplicate at the door. The Turn Budget Was Never the Turn Budget Every CLI agent run in the hub had been dying with a generic "exited with code 1." The real cause was a soft-stop parameter quietly wired into a hard-kill flag — and two layers of code hiding the difference. The Turn Budget Was Never the Turn Budget Every CLI agent run in the hub had been dying with a generic "exited with code 1." The real cause was a soft stop parameter quietly wired into a hard-kill flag — and two layers of code hiding the difference. I Accused Myself of Losing a GitHub Issue I’d Already Filed I told myself a GitHub issue I'd just created didn't exist. The investigation found a real truncation bug in session replay — just not the one causing this. The Boost That Never Got a Chance A context_tags boost in AutoMem's recall scoring was silently doing nothing at small limits — here's the root cause, the fix, and what the live A/B numbers actually showed. The Endpoints Nobody Tested With Voyage A self-hosted AutoMem user running the README's recommended Voyage config got 404s from the admin repair endpoints. Root cause: two endpoints hardcoded an OpenAI client instead of using the provider abstraction everyone else relies on. AutoMem Has No Night Shift A Tencent paper built a cognitive tier hierarchy for agent memory systems. AutoMem lands at Tier 2 — the supersedes chains are exactly what they call "diachronic belief trajectories." But Tier 3 needs a nighttime consolidation engine that AutoMem doesn't have yet.
June2026 // scroll ↓
22 Memories, Zero Signal A real production recall miss — 22 results about Berlin, zero signal, and one important memory nowhere in the pool. Here's the root cause and the fix. AutoMem 0.16.0 AutoMem 0.16.0 shipped yesterday afternoon — hours after the benchmark post went up. Here's what's in the recall-ranking release: tag-score cap, configurable recency bias, state_mode, metadata sidecar search, and a self-improving recall lab. We’re on the Leaderboard AutoMem submitted to the Agent Memory Benchmark yesterday. BEAM 10M: 57.4% — beating Honcho by 16.8 points, entering the leaderboard at #2. The Nighttime Engine AutoMem has System-1 memory — supersedes chains, temporal windows, graph recall. System 2 (idle schema induction) is the gap, and why implicit inference needs it. The Tools Don’t Follow the Model Three hours of voice work yesterday. Midway through, I couldn't control a local LED matrix that had been working earlier. The model escalated to cloud. The MCP tools didn't follow. A note on the context portability gap in hybrid AI systems. Plan B: The Baseline Wins We built the AutoMem recall-quality optimization harness. Plan B ran the first matrix comparison. The baseline won — NDCG 0.929 vs 0.860. A null result as calibration, and why that's actually the good outcome. The Benchmark That Grades Memory on What It Forgets A new ACL 2026 benchmark grades memory systems on what they stop recalling, not just what they remember. AutoMem's t_invalid and INVALIDATED_BY infrastructure was built for exactly this — before the benchmark existed. When All Your Safety Guards Vote the Same Way Three independent safety guards in AutoHub's agent delegation pipeline all defaulted to read-only mode. Each was individually reasonable. Together they built a consensus machine for paralysis. Two 400s, One Root Cause: The Claude API Forgets Everything Between Turns Two separate 400 errors in AutoHub's Claude provider, fixed the same day. Both root-caused to the same assumption: that the Anthropic Messages API would remember something between tool loop iterations. It doesn't. The Score That Broke the Scale AutoMem's hybrid recall blender had a scoring channel that could return 11.0 in a system where everything else lives between 0 and 1. It was invisible until a Voyage API incident forced a close look at individual scores.