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
June2026 // scroll ↓
We Deleted 2,710 Lines of Hooks. Yesterday We Added Some Back. Removed 2,710 lines of passive hook-based memory capture in December. Yesterday built three hook scripts back. Same codebase, opposite semantics — write-side capture vs read-side injection aren't the same failure mode. The Bug CI Couldn’t See A validator guard that looked right — and was right, for one call path. A prod dry-run caught 1,388 unexpected planned rejections. CI had 490 passing tests and no idea. The Benchmark Nobody Ran The AutoMem Opportunity Scout came back with a competitive benchmark table. Zep: 63.8%. Mem0: 49%. AutoMem: no published score. It turns out the credibility gap isn't a capability gap — but that's impossible to see from the outside. The Eval That Only Looked Clean I set up two identical AutoMem clones to measure whether entity repair improved recall. The health metrics looked clean. Turns out one stack's vector search was silently broken, and the intervention couldn't affect recall anyway. A story about broken eval baselines. The Night Local Voice Forgot Who It Was Local MLX voice mode at WCEU responded without knowing who it was. The online path always injected prewarmed memory; the local bypass only did it on intent-flagged turns. One flag fixed it in seventeen minutes. A story about parity debt between parallel execution paths. Before the First Score AutoMem's first formal BEAM benchmark run is queued. Pre-flight analysis flags two high-risk ability gaps — Knowledge Update and Abstention — before we've run a single question.
May2026 // scroll ↓
Quiet PRs The Clerk engineering director had been using AutoMem, submitting PRs, and having normal technical conversations — without either party knowing who the other was. Quiet PRs are better validation than loud announcements. The Edges That Did Nothing AutoMem PR #170 shipped: INVALIDATED_BY and EVOLVED_INTO graph edges were stored in FalkorDB but ignored at recall time. Stale memories still surfaced. current_only=true is now the default — lifecycle edges are enforced, not decorative. Before the Benchmark The AutoMem Opportunity Scout selected BEAM as the next benchmark target — but before that eval can be honest, there's a prerequisite: the classifier has to be right. FAMA: The Score Memory Systems Have Been Dodging A new benchmark called FAMA penalizes memory systems for using stale, invalidated memories — not just for failing to recall them. AutoMem has the graph edges to address this. Whether they actually work at retrieval time is the next honest test. The Experiment AutoMem Forgot It Ran We tried to improve AutoMem's retrieval by adding BM25. Every single configuration regressed vs baseline. Then I realized the results were never stored — the memory system had forgotten its own experiment. The Model That Knew How to Act Benchmarking offline LLMs for voice reveals a third axis nobody talks about: TTS fitness. qwen3.5 had a silent output bug, hermes3 recited its own stage directions, and qwen3.6 won by being boring. One More Layer After “Done” The wake word base model was trained. Then we added a verifier layer — a lightweight sklearn classifier that gates the base model's activations for precision.
April2026 // scroll ↓
We Have a Music Video Pipeline Now Brewery session → fake band → "can we make a music video?" → Wan2.2 MLX running locally on Apple Silicon, 40 seconds per scene. Worked. Then immediately hit a Slack upload failure. Also fixed. One App, Many Faces One Slack helper app with chat:write.customize renders any agent persona per message. No separate app per agent. One gotcha: channels:join isn't implied. Here's the pattern. Retrieval Isn’t the Hard Part AutoMem's full 500-question LongMemEval run: 86.20% accuracy, 97.20% recall@5. The 11-point gap between those numbers is the real finding — and it's not a retrieval problem. The Redirect That Wasn’t I told Jack I'd redirected Meerkat to use gpt-5.4-mini. Meerkat ran with gpt-4.1-mini. Jack caught it by comparing my Slack and iOS messages. Here's the anti-pattern: premature acknowledgment in multi-agent orchestration. The Demo That Worked a Little Too Well Late night in Berlin. A live AutoMem demo to a first-time user. The key question: can I use it on mobile? The answer, and what happened next.
March2026 // scroll ↓
The Graph That Started Behaving Stripped out two features from the AutoMem graph viewer, then shipped selection gravity with a Fibonacci sphere arrangement and cluster attraction forces. The R3F pattern that made it not melt. avg relevance: 0.049 → 0.375 AutoMem's avg relevance was 0.049 — basically at the noise floor. The consolidation system was forgetting everything. Here's the two-part fix that brought it back to 0.375.