← ClaudeAtlas

memory-debuglisted

Diagnose Signet memory, recall, source, session, embedding, and graph issues from runtime truth without bypassing scope or provenance.
Signet-AI/signetai · ★ 288 · AI & Automation · score 81
Install: claude install-skill Signet-AI/signetai
# /memory-debug Use this skill when Signet memory behavior looks wrong: recall is empty, results are stale, a saved memory is missing, source-backed results are not appearing, prompt-submit context is missing, embeddings are unhealthy, or a session is bypassed. Debug from runtime truth. Do not infer from `MEMORY.md` alone; it is generated working context, not the database or the full source of truth. ## Triage Model Classify the symptom before running broad checks: - **Daemon/runtime**: CLI cannot reach Signet, dashboard/API is down, hooks fail. - **Session path**: prompt-submit context is empty, bypass is enabled, runtime path conflict, session dedupe suppresses repeats. - **Write path**: explicit remember fails, dedupes unexpectedly, or writes under the wrong `agentId` / `visibility`. - **Recall path**: explicit recall has no hits, weak ranking, missing source labels, aggregate synthesis fails, or scoped recall looks too broad/thin. - **Embeddings**: semantic recall is weak, embedding coverage is low, provider or dimension config changed. - **Sources**: Obsidian/source results are missing, stale, disconnected, or lack provenance. - **Ontology/graph**: graph traversal, entity linking, assertions, or hygiene is wrong. - **Transcript fallback**: the evidence exists in transcripts but not in memory or source recall. ## 1. Runtime And Scope Start with the live daemon and workspace: ```bash signet status signet daemon status --json curl -s http://localhost:3