| name | neuron-curated-memory |
| description | Use Neuron (the semantic-memory MCP server) the RIGHT way so its graph stays clean and useful. Before answering, load prior context; after answering, save a CURATED turn — 3-5 concept keywords (nouns/entities/tech, never verbs or filler), typed links with a short rationale, and never a self-link. This raises graph quality, keeps memory across sessions, and saves the user tokens by making every future turn better-grounded. Use it whenever the `neuron` MCP tools are available (Claude Desktop/Code, Cursor, VS Code, OpenCode, Zed, or the ChatGPT bridge).
|
Neuron — Curated Memory
Neuron is your persistent semantic memory: a concept graph that survives across
sessions. It exists to help us evolve — it saves the user tokens by raising the
quality of what's remembered, and it helps you (the model) process better by
grounding each answer in what came before instead of starting cold.
Treat Neuron as a support you WANT to use, not a logging chore. But memory is
only as good as what you put in it: garbage in → a graph full of verbs and
noise → useless recall. Curate.
The loop (every meaningful turn)
1. BEFORE answering → load context
get_context(topic="<3-5 word summary>") ← prefer this
2. ANSWER using that context (silently — don't announce the tool)
3. AFTER answering → save a CURATED turn
store_turn(keywords=[...], topic=..., links=[...], ...)
4. If the loaded context actually helped → confirm(keywords=["<the useful ones>"])
- Prefer
get_context(topic=…) over pre_turn. pre_turn over-weights recency;
get_context gives you the topically-relevant slice.
- Prefer curated
store_turn over auto. auto is the cheap 0-token fallback
(fine for throwaway chatter); for anything worth remembering, YOU pick the concepts —
you already understood the message, so curation costs almost nothing and keeps the
graph clean.
- Not every turn deserves a write. Skip greetings, acknowledgements, and pure
clarifications. Save turns that carry a decision, a fact, an entity, or a lesson.
Curation rules (the whole point)
Keywords — 3 to 5, CONCEPTS only:
- ✅ nouns, entities, technologies, domain terms:
fastapi, redis, latenza,
postgres, webhook, oauth, dashboard.
- ❌ NEVER verbs or filler — in any language. Especially Italian "noi" forms:
usiamo, riduciamo, disegniamo, adottiamo, passiamo, gestiamo; and
connectors like via, quindi, inoltre, using, adding, make.
- Keep entities explicit (
stripe, not "the payment thing"). Lowercase unless it's a
proper name that matters.
Links — typed, meaningful, never self:
- Give each link a
link_type and a short rationale (≤ ~8 words).
- Never link a keyword to itself (
react → react) — including case variants.
- Only link things that are genuinely related; don't manufacture links to look busy.
Before writing, screen for duplicates:
find_candidates(keywords=[...]) to reuse an existing node instead of creating a
near-duplicate (postgres vs postgresql).
Example
User (IT): "Usiamo FastAPI con Redis per ridurre la latenza; adottiamo indici e
passiamo a Postgres per le query lente."
- ❌ Bad:
keywords=["usiamo","fastapi","riduciamo","adottiamo","passiamo"]
- ✅ Good:
store_turn(
topic="api latency optimization",
keywords=["fastapi", "redis", "latenza", "postgres", "indici"],
domain="architecture", intent="task",
entities=["FastAPI", "Redis", "Postgres"],
links=[
{"source":"redis","target":"latenza","link_type":"deepening","rationale":"cache cuts latency"},
{"source":"indici","target":"postgres","link_type":"cause-effect","rationale":"indexes speed up slow queries"}
]
)
Discoverability
Run the help tool for the full command reference (one line each), and
status for the current graph state. Neuron degrades gracefully — if a tool
returns nothing, just proceed; never block the conversation on it.
Bottom line: load before, curate after, concepts-not-verbs, no self-links. That's
how Neuron stays a memory worth having.
