Skip to content

GitHub Copilot Agent Skills (HydraMem)

HydraMem ships six Copilot skills in /home/runner/work/hydramem/hydramem/.github/skills/ so agent-mode workflows can use a predictable pattern to ingest knowledge, recover context, and maintain graph quality.

Skill map

Skill Purpose Main MCP tools Context flow
hydramem-ingest Ingest Markdown files/directories ingest_markdown, ingest_directory_tool Adds new searchable chunks/entities.
hydramem-ingest-smart Agent-led semantic ingestion ingest_prechunked, submit_session_extraction Injects high-quality structured context into memory.
hydramem-query Direct factual lookup with citations priming_context_tool Recovers top chunks + neighbours, then injects them into answer prompt.
hydramem-reason Multi-hop reasoning across graph hydra_search_tool Recovers expanded graph context for cross-document reasoning.
hydramem-link Manual relation curation create_relation, delete_relation, check_conflict_tool Improves graph structure so future retrieval is better grounded.
hydramem-garden Autonomous maintenance cycle get_garden_status_tool, run_night_gardener, train_gnn_tool Consolidates and prunes context after ingestion bursts.

Context injection and retrieval patterns

1) Inject context into HydraMem

Use these when the goal is to store new knowledge:

  • hydramem-ingest for standard Markdown indexing.
  • hydramem-ingest-smart when the agent can read and chunk the content itself and submit entities/relations directly.
  • hydramem-link when you need to manually add/remove explicit graph edges.

2) Recover context for answers/search

Use these when the goal is to answer with grounded memory:

  • hydramem-query for direct questions, where priming_context_tool retrieves focused evidence blocks for citation-backed answers.
  • hydramem-reason for multi-hop or causal questions, where hydra_search_tool expands through graph links to build richer context.

3) Keep context quality high over time

  • hydramem-garden runs inference + verification + pruning after heavy ingestion or on periodic maintenance windows.
  1. Ingest docs/notes (hydramem-ingest or hydramem-ingest-smart).
  2. Ask factual questions (hydramem-query) or deep relationship questions (hydramem-reason).
  3. Curate critical edges manually when needed (hydramem-link).
  4. Run maintenance (hydramem-garden) after significant updates.

This keeps retrieval grounded while steadily improving long-term context quality.