Skip to content

query.md

Latest commit

 

History

History
58 lines (41 loc) · 2.9 KB

query.md

File metadata and controls

58 lines (41 loc) · 2.9 KB

memory query

memory query asks a project-specific question against curated Memory Layer data.

Use it when you want a direct answer from durable project memory instead of browsing the TUI. The answer is synthesized from the returned memories and includes citation numbers that map back to the ranked results.

Common Usage

memory query --project memory --question "How does resume work?"
memory query --project memory --question "What changed recently?" --type plan
memory query --project memory --question "What was actually implemented for the watcher manager?" --type implementation
memory query --project memory --question "What code was refactored recently?" --type refactor
memory query --project memory --question "How is PostgreSQL setup documented?" --type documentation
memory query --project memory --question "What are the watcher health states?" --tag watcher

Useful Flags

  • --type restricts retrieval to one or more memory types
  • --tag restricts retrieval to one or more tags
  • --limit caps how many memories are considered before answer synthesis
  • --min-confidence filters out weaker memories
  • --json returns the full result payload

Output

The default text output shows:

  • the synthesized answer
  • confidence and whether evidence was sufficient
  • the answer-generation method (llm, deterministic, or fallback)
  • cited memory numbers, matching the ranked result list
  • retrieval diagnostics and provenance highlights

If LLM answering is configured, the backend asks the model to answer using only the returned memories. Supported LLM providers are openai_compatible and ollama; Ollama uses http://127.0.0.1:11434/v1 and no API key by default. If the model is unavailable or returns invalid citations, Memory Layer falls back to deterministic summary synthesis and reports the fallback reason.

Graph-Aware Retrieval

When a project has a completed memory graph extract run, memory query automatically uses the latest completed code graph as an additional retrieval signal. No extra flag is needed.

Graph retrieval is additive:

  • lexical and semantic matches still determine the baseline result set
  • graph matches can add memories whose file provenance points at matching symbols or one-hop related symbols
  • graph boosts are capped so code graph hints cannot overwhelm strong memory matches
  • graph connections are shown as explanations, not as standalone answer citations

The JSON output includes graph diagnostics such as graph_status, graph_candidates, graph_augmented_candidates, and graph_duration_ms. Individual results can include graph_connections describing the file, symbol, edge, neighbor symbol, reason, and score boost that affected ranking.

Related Docs