chl-dynamic-retrieval

name	chl-dynamic-retrieval
description	Use when the task is to dynamically recover, inspect, or synthesize information from the CHL Memory MCP by choosing the right tool or resource at runtime.

CHL Dynamic Retrieval

Use this skill when you need to recover information from chl-memory without hard-coding a single tool path.

Goal

Choose the narrowest CHL source that answers the request:

• exact read-only state: use MCP resources
• semantically related memories: use chl_recall
• synthesized answer from memory: use chl_infer
• debugging or auditing: use chl_state, chl_entries, chl_journal, or chl_bucket_stats

Retrieval Strategy

1. Classify the request:
   • "What do we already know?" -> chl_recall
   • "What is the best answer?" -> chl_infer
   • "Show me the current data/state" -> resource read or snapshot tool
   • "Why is retrieval failing?" -> diagnostic tools
2. Start narrow:
   • keep queries short
   • begin with topK: 3 or topK: 5
   • prefer one focused query over many broad ones
3. Refine only if needed:
   • broaden the query
   • raise topK
   • inspect the lexicon if the user is using domain-specific terms

Tool Selection

• Use chl_recall when you want similar memories, supporting context, or candidates to inspect.
• Use chl_infer when the answer should be synthesized from retrieved memories.
• Use chl_snapshot for a compact current view.
• Use chl_profile when memory behavior depends on the active profile.
• Use chl_lexicon when terminology, concepts, or phrase mappings matter.
• Use chl_state for a full snapshot of the system.
• Use chl_entries when you need every stored entry.
• Use chl_journal when you need the mutation history.
• Use chl_bucket_stats when you suspect retrieval quality or indexing issues.

Resource-First Reads

If you already know the exact data you need, prefer resources:

• chl://memory
• chl://profile
• chl://state
• chl://entries
• chl://journal
• chl://lexicon
• chl://lexicon.concepts
• chl://lexicon.phrases
• chl://lexicon.tsv

Use resource reads when the answer is mostly inspection rather than search.

Dynamic Heuristics

• If the user names a specific memory or phrase, search with that wording first.
• If the user asks a broad question, start with chl_infer.
• If the first result set is noisy, inspect the lexicon and retry with more precise terms.
• If the request is about persistence, exports, or recovery, switch to backup tools instead of recall.

Guardrails

• Do not call several broad tools in parallel unless the request truly needs comparison.
• Do not guess missing context when the MCP can supply it.
• Do not clear or overwrite memory unless the user explicitly asked for that.