hermes-history-ingest

Ingests Hermes agent history into an Obsidian wiki, mining past Hermes sessions and memories for durable knowledge and distilling it into wiki pages. It triggers when a user wants to process their Hermes history, import their ~/.hermes folder, or extract insights from previous Hermes conversations, and can be invoked directly or through the wiki-history-ingest router. The emphasis is on capturing lasting knowledge rather than operational telemetry.

Before running, it resolves configuration via the Config Resolution Protocol in llm-wiki/SKILL.md (yielding OBSIDIAN_VAULT_PATH and HERMES_HISTORY_PATH, defaulting to ~/.hermes), then reads the vault's .manifest.json and index.md. It supports two ingest modes: Append (default), which processes only files absent from the manifest or modified since their ingested_at timestamp, and Full, which reprocesses everything after a rebuild or on explicit request. The Hermes data layout under ~/.hermes includes memories (highest-signal curated knowledge in markdown or JSON), sessions (turn-by-turn JSONL transcripts that are rich but noisy), config.yaml (metadata only), and a .hub directory that is skipped along with the skills directory.

The workflow surveys sources and computes a delta, parses memories first, then parses session JSONL safely, clusters extracted knowledge by stable topic, and distills it into wiki pages routed by convention (projects, concepts, skills, entities, synthesis). Writing rules require summary frontmatter, confidence and lifecycle fields, and provenance markers such as extracted, inferred, and ambiguous. A critical privacy filter mandates removing API keys, tokens, passwords, and credentials, redacting private identifiers, and summarizing rather than quoting transcripts verbatim. Finally it updates the manifest, index.md, log.md, and hot.md, and optionally refreshes the QMD search index via qmd update or qmd embed when a collection is configured.