Use when the user is working with StudioMeyer Memory and you need to pick the right tool. Covers when to use search vs entity tools vs learn vs decide, how to structure queries, and how to avoid duplicate writes.

StudioMeyer Memory Workflow

This skill teaches the right tool choice for any memory operation. The Memory server has 53 tools — this is the decision tree for the 10 you will use most.

Core principle

Memory has four types of content:

Learnings — facts, insights, mistakes, patterns. Append-only with a confidence score. Use nex_learn.
Decisions — tracked choices with rationale and outcome. Use nex_decide.
Entities — typed nodes in the Knowledge Graph (people, projects, tools, services). Use nex_entity_*.
Sessions — conversation containers with start/end markers. Use nex_session_*.

Search spans all four with nex_search. It is the default for "do we know anything about X?" questions.

Decision tree

User asks about something we might already know

→ nex_search with a natural-language query. Let expand: true (the default) handle synonym expansion. If the query is aggregation-heavy ("how many", "all the", "summarize"), set agentic: true.

User wants to save a fact or insight

→ nex_learn with the right category:

mistake — something that went wrong
pattern — recurring technique
insight — non-obvious realization
research — findings from investigation
architecture — design decision
workflow — process
security — security-relevant fact

User makes a choice with a reason

→ nex_decide with content, rationale, outcome (if already known). Decisions can be linked with nex_decide follow-ups later.

User mentions a person, project, tool, or service

→ nex_entity_search first to check for duplicates. If not found, nex_entity_create. Then nex_entity_observe to add facts and nex_entity_relate to link it to other entities.

User asks "what's happening" or "what's next"

→ nex_sprint (current tasks) plus nex_proactive (stale items, open decisions, knowledge gaps).

User wants a deeper read

nex_synthesize — generate a guide from related learnings
nex_reflect — extract meta-insights from recent work
nex_insights — pre-computed insights from the last session

Anti-patterns

Dumping entire conversations into nex_learn. The server has a Gatekeeper that will reject low-signal saves. Distill first.
Calling nex_search with a single word like "crm". Too broad, too many results. Use 3-6 meaningful words.
Creating a new entity every time. Always nex_entity_search first. Merge with nex_entity_merge if a duplicate slipped through.
Starting a session you will not close. Call nex_session_end at the end of real work. Orphan sessions pollute analytics.
Ignoring nex_proactive at session start. It surfaces stale items the user forgot. Cost is near-zero.

Hard limits that matter

nex_search default limit is 20. Set limit: 5 for quick checks, keep 20 for real investigations, cap at 100.
Cross-agent search is default-on. If you want to filter by a specific agent, pass agentId: "name".
recencyWeight defaults to 0.3. For "what did we decide recently" set it to 0.7. For "is this a known pattern" set it to 0.1.

Example recipes

"Did we ever solve the Docker memory issue?"

nex_search({ query: "docker memory issue container crash", recencyWeight: 0.2, limit: 10 })

"Remember that we decided to use Supabase Frankfurt for all new SaaS projects."

nex_decide({
  content: "Use Supabase EU Frankfurt for all new MCP SaaS projects",
  rationale: "GDPR compliance, latency for EU users, existing StudioMeyer PRO org",
  confidence: 0.9
})

"Find everything about the GEO product."

nex_entity_search({ query: "GEO", types: ["project", "service"] })
// then for the matching entity:
nex_entity_open({ entityId: "..." })
nex_entity_graph({ entityId: "...", depth: 2 })

memory-workflow