| name | neuron-usage |
| description | Correct-usage playbook for the Neuron semantic memory MCP (neuron5 tools). Use when storing or recalling memories with Neuron, when deciding what to pass to pre_turn/store_turn, when the user says "remember this", "salva in memoria", "cosa ricordi di", "usa neuron", when curating keywords/links for the knowledge graph, or when Neuron tools return errors or unexpected results. |
Neuron — correct usage playbook
Neuron is a persistent semantic memory: a knowledge graph of concept nodes and
typed links, shared across sessions (and possibly across a team via Turso
Cloud). Every bad write pollutes it for everyone. Follow these rules exactly.
The per-turn loop
mcp__neuron5__pre_turn(topic, keywords) BEFORE replying on any
substantive turn. Fold the returned context silently into the reply —
never announce the call or quote its raw output.- Reply to the user.
mcp__neuron5__store_turn(topic, keywords, links, domain, intent, sentiment) AFTER replying, to persist only what is NEW.
Skip 1 and 3 on procedural turns (acknowledgements, thanks, yes/no) and when
the graph is empty. Call mcp__neuron5__help once per session if unsure of
the command list.
Curating store_turn input (the part models get wrong)
- topic: 3-5 words, descriptive, no punctuation soup.
- keywords: 3-5 singular concept NOUNS — entities, technologies, ideas.
- GOOD:
"retry backoff", "install manifest", "spreading activation"
- BAD: verbs (
"implement", "fix"), sentences, file paths, whole
phrases, filler ("use", "make"), duplicated casing variants.
- Allowed characters: letters, numbers, spaces,
-_.:+ (the server
rejects anything else).
- links: only between keywords OF THIS TURN (or a current keyword and a
well-known previous one). Always typed
(
cause-effect | analogy | evolution | contrast | deepening | instance-of),
never a self-link, weight honest: tangential if unsure, strong only
for load-bearing relations.
- Before minting a concept that may already exist: check with
mcp__neuron5__vector_search(query) or mcp__neuron5__find_candidates and
reuse the existing keyword instead of creating a near-duplicate
("db layer" vs "database layer").
- NEVER store secrets, tokens, passwords or personal data as concepts.
Reinforcement and maintenance
mcp__neuron5__confirm(keywords) when context returned by pre_turn was
actually useful — it boosts salience of the right nodes.mcp__neuron5__dedup / mcp__neuron5__merge for duplicates;
mcp__neuron5__consolidate for periodic cleanup (do NOT run consolidate
ad-hoc on a SHARED cloud store — coordinate with the team first).mcp__neuron5__switch_context(path) to work in a separate context
(e.g. java/spring); one context per feature/area is the recommended
team layout.
Common mistakes to refuse
| Mistake | Do instead |
|---|
| store_turn on "ok thanks" | Skip the loop on procedural turns |
| keywords = "fixed the installer bug" | "installer", "registration engine" |
| link a keyword to itself | Drop the link |
| announce "let me check my memory" | Use pre_turn silently |
| paste pre_turn output verbatim in the reply | Weave the context in naturally |
| store an API token as a node | Never — refuse and warn |
| re-calling pre_turn multiple times per turn | Once per user turn |
Full documentation
Call mcp__neuron5__skill with name='playbook' for the complete
playbook served by the server itself, or mcp__neuron5__help for the
one-line command index.
