Memuron lets AI agents remember.
Memory that decides, merges, links, and evolves.
For teams building AI agents that can't afford to hallucinate — in fintech, healthcare, and other high-stakes systems. Memuron resolves contradictions at write-time, so your agent never has to guess which memory is true.
Agents have amnesiaVector DBs make it worse.
Tell an agent “I like blue” on Monday and “actually, I prefer red”on Tuesday. A plain vector store returns both — a contradiction it can't resolve.
Memuron decides at write time: is this a new fact, or an update to one you already hold? Then it collapses the two into one evolved memory.
An LLM pipeline that shapes memory before it commits.
Every ingest runs through the Guardian: it retrieves candidates with hybrid search, decides create-vs-update, draws semantic links, and commits typed events to an append-only ledger.
Raw content arrives from the SDK or API and is accepted as an immutable payload.
A memory system, not a vector bucket.
Agent-repairable memory graph
live · hover a nodeMemuron stores every relationship as a first-class memory_link, including self-loops and parallel edges agents can inspect later.
Hybrid graph search
Vector, lexical, and link-description matches surface memories plus the edge that made them relevant.
Context assembly
Prompt-ready blocks with citations, graph links, breadcrumbs, and a hard budget.
URL and folder ingest
Fetch pages, sync local folders, and keep files mapped to stable graph nodes.
Stable source identity
Attach custom_id, session, thread, source ID, and URL to every memory or document.
Multi-tenant scoping
Cryptographic tenant_id boundaries plus spaces for project, team, and user context.
••••
••••
MCP native
Connect Cursor over an HTTP stream, authenticated with Clerk OAuth — no key wrangling.
Semantic traversal
Don't just match text. Ask a question and walk the edges that answer it.
Pipe semantic search through your graph.
One query, three stages. Toggle each below and watch the set narrow — from a fuzzy semantic match, down the graph's edges with traverse, to a final grep.
Which rollback decisions were tied to recent deploy incidents?
First, find memories that mean “deploy decisions.”
Semantic search reads for meaning, not keywords — so it surfaces anything about shipping, even without those exact words.
Memuron is the chassis. Arthaanu is the engine.
Memuron handles the product — tenancy, ingest jobs, the Guardian. The horsepower comes from Arthaanu, an open-source semantic engine you can inspect, self-host and swap: an append-only ledger, deterministic encoders, rebuildable projections, and the retrieval math underneath.
the append-only source of truth — objects, encodings, relations and traces
“v2.3 deploy was rolled back.”
0.88, 0.31,
-0.17, … ]