David F GliddenandClaude Opus 4.8
3f9a89b00c
Basic Memory v0.21.6 first sync over the live memory dir (steward-authorized live-dir trial, Option A 2026-06-06): adds permalink: to frontmatter, refolds long YAML description lines, strips final newlines. Bodies untouched — verified via full diff classification. From this commit forward, any diff in claude/memory shows only what Basic Memory or the session writes. Trial design: MemPalace untouched as incumbent; git status check on this dir at every wrap; end-of-day evaluation (recall quality, sync robustness, rebuild-from-files, malformed-file behavior). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
1.0 KiB
1.0 KiB
name, description, type, permalink
| name | description | type | permalink |
|---|---|---|---|
| cm-david:v2 is noisy — do not use as BM_LOCAL_MODEL | Fine-tuned model from v0.48.0 era produces garbage on structured slot tasks — retrain on clean v0.61.0 pairs | feedback | claude-memory/feedback-cm-david-v2-noisy |
cm-david:v2 (qwen2.5 LoRA, 14,199 pairs) was tested on classification and entity resolution tasks. It produces rambling, hallucinated output instead of the terse structured responses the slots expect. The training pairs were generated during the v0.48.0 era when L1 had reliability bugs (A-D).
Why: Noisy pipeline → noisy pairs → noisy model. The teacher (Haiku) corrections were likely valid but the training context was corrupted.
How to apply: Use qwen3.5:4b as BM_LOCAL_MODEL until cm-david:v3 is trained on clean v0.61.0 pairs. Purge old training pairs (~/.capablemind/data/mindfabric-00/training/*.jsonl) before generating new ones. The cm-keystone-ranking models (3 versions) in Ollama are also suspect — trained on the same era's data.