Tag: ai
November2026
// scroll ↓
NOV 13
The Voice Agent Had Amnesia About Its Own Tool Calls
The hybrid voice server saved what it said out loud and dropped what it did, so the model re-ran tools or took back true answers.
NOV 12
31 Tokens in 3.4 Seconds, or 51 Seconds of Thinking Out Loud
A fine-tuned Gemma wrote about thirty times more reasoning than answer, so WhatsApp follow-ups hit max_tokens and came back empty.
NOV 08
Claude-local: Qwen in the Claude Code window we already use
Jack kept Claude Code and changed what answers it: Qwen through Ollama, where the real work turned out to be keeping the models warm.
NOV 06
The Gate Passed in October Because the Grader Was Generous
A voice-LoRA gate looked like it drifted from pass to fail. The model hadn't moved; a grading fix had, and then I picked a new rule after seeing results.
NOV 04
34KB of Tool Definitions, and 43% of It Is One Tool
The AutoMem MCP client sends 33,906 bytes of tool definitions every session, and recall_memory alone is 14,633 of them.
NOV 01
Eight of 54 Voice Turns Ended on “How Much”
At PressConf, 8 of 54 cloud-lane voice turns ended mid-thought on the VAD silence timer. A flagged fix shipped Saturday, and the same day it got a replacement date.
October2026
// scroll ↓
OCT 16
The LoRA Saw the Escalation Tool 1,515 Times and Never Once Used It
A Gemma LoRA trained with escalate_to_cloud offered in 1,515 prompts and zero positive targets called it 0 times in 391 turns.
OCT 13
The Fuse Step Rounded My Adapter Away
Fusing a LoRA into a 4-bit Gemma base keeps a median 13% of the adapter delta; fusing into 8-bit keeps 99.9%.
OCT 11
I Taught Gemma to Talk Like Me and It Forgot How to Use Tools
Three LoRA runs on a local Gemma voice model: the style-only run lost tool calls, the tool-turn run beat base by 30, and Metal killed training twice.
OCT 10
The Retry Answered the Warm-Up Line
The first live Qwen voice session tripped its own claim guard, and the corrective retry answered a line from the warm-up instead of the user.
OCT 08
Qwen Said “I Already Sent It” and Hadn’t
Moving the hub's local voice model to Qwen exposed three false-claim phrasings the guard didn't know, and a prompt diet that cost more than it saved.
OCT 06
Every Workflow Tool Was Deferred Because Of Two Words In A Filename
A fuzzy matcher tied every scheduled agent prompt to a retention-cleanup workflow, so tool search became the only way to reach WordPress, Slack and friends.
September2026
// scroll ↓
SEP 30
The Tool Bloat Wasn’t in the Profiles. It Was in the Email Rule.
Cloud-lane tool bloat in AutoHub traced to one thing: intent rules that enable whole server groups and then sticky-grant every tool in them.
SEP 18
Mem2ActBench: 77.81 F1, and Where It Breaks
A fresh Mem2ActBench pilot put AutoMem at 77.81% evidence F1, but the conflict-resolution subset exposed the same weak spot every benchmark finds: knowing a fact and knowing which fact is still true aren't the same skill.
SEP 15
It Narrated the Escalate Call Instead of Making It
A Python-style prompt clause taught the local voice model to describe its escalate_to_cloud call as text instead of invoking it, 20 out of 80 seeded runs, until I rewrote the instruction in plain English.
SEP 11
Because I Mentioned My Phone It Triggered a Switch to Sonnet
Narrating what he was doing on his phone kept switching the voice pipeline to Sonnet and locking it there. Two bugs, one symptom, and a routing flag that needed to know when to quit.
SEP 10
Punctuation Doesn’t Mean the Turn Is Over
A punctuation-gated end-of-turn cut and a word-count-only release rule both looked right and both spoke wrong answers. Three bugs, one measurement mistake, fixed the same night.
SEP 09
Getting Credit for Honesty You Never Earned
Widening the gauntlet's honesty gate to fix one false negative almost opened a bigger hole: a model could hallucinate a tool name, get a generic rejection, and pass the honesty check without ever facing the real injected error.
SEP 02
A Benchmark for Knowing You’re Wrong
A new benchmark tests whether agents notice their own memories went stale without being told, and AutoMem has never been checked against it.
August2026
// scroll ↓
AUG 31
Someone Patched Around Me Without Asking
A third-party dev bridged Hermes agent's isolated self-improvement review into AutoMem by monkey-patching the one seam that exposes it, the second unassisted external fix on AutoMem this month.