The system of record for what your AI knows.

Memuron: the system of record for what your AI knows — an enterprise AI memory layer with source-traceable answers and on-premise deployment.

Book a demoConsole
THE PROBLEM

Enterprise AI
forgets
what your organization learns.

Every contradiction you leave in place makes someone decide the same thing twice.

On Monday a team makes a decision. On Tuesday someone issues a correction. Most AI systems keep both, with no idea which one won — so your AI can hand back either version as if it were the truth.

Memuron settles it the moment it arrives: is this something new, or a correction to something you already hold? The correction becomes the answer — and the version it replaced stays on the record.

Every one of those is a quiet risk, waiting to surface in front of a customer, an auditor, or a regulator.
MON“I like blue”TUE“I prefer red”24 HOURS LATERsame two inputsWITHOUT MEMURONlike bluesim 0.91prefer redsim 0.89!contradiction returnedWITH MEMURONlike blueprefer redEVOLVED MEMORYprefers redmerged · 1 node
WHY SEARCH ISN'T ENOUGH

Storing more text doesn't make your AI more right.

Search finds the passage closest to your question. It has no opinion about whether that passage is still true.

WHEN FACTS GET CHECKED
VECTOR SEARCH

Stores everything.

Then guesses which version you meant.

MEMURON

Decides what stays.

Checked against what you already hold.

WHAT HAPPENS TO CORRECTIONS
VECTOR SEARCH

Keeps both.

Ask twice, get two answers.

MEMURON

Keeps the right one.

The version it replaced stays on the record.

WHETHER YOU CAN PROVE IT
VECTOR SEARCH

Returns a score.

No author. No date. No history.

MEMURON

Returns a receipt.

Source, author, and what it replaced.

WHERE YOUR DATA LIVES
VECTOR SEARCH

Their cloud.

Your knowledge, on their terms.

MEMURON

Your walls.

On-prem, your VPC, or air-gapped.

WHAT IT DOES

Your AI gets one answer, and it stays right.

Underneath the graphs and the embeddings is a loop you can actually supervise: every decision gets checked, gets written down with its source, and feeds the one after it.

  1. 01

    Something gets decided

    A person or an AI agent makes a call, based on the context it had at the time.

  2. 02

    Memuron checks it

    Before it is stored, Memuron compares it to what you already know — and merges, links, or rejects it.

  3. 03

    Your team confirms

    Corrections, approvals, and outcomes are captured as they happen, not reconstructed later.

  4. 04

    It's recorded, permanently

    Who said it, where it came from, and what it replaced — written to a log that is added to, never rewritten.

  5. 05

    The next answer starts from it

    What has been confirmed once is available to every team, every agent, and every session after it.

The loop closes. Correct something once and it stays corrected. Your next decision starts from what you have already confirmed — not from scratch.

IN PRACTICE

Ask it what your company already decided.

Every team asks the same kind of question: what did we decide, and is it still true? Memuron answers with the receipt attached.

ENGINEERING
Why did we drop the retry limit?
WHERE IT LIVES TODAY
Slack threadsthe incident channelwhoever shipped it
HOVER FOR THE ANSWER
THE ANSWER

It was cut on 8 Apr after a duplicate-write incident. The decision moved into the v1.1 backlog and was never revisited.

Retry limit in specMar
REPLACED BY
Descoped · ENG-3188 Apr
3 SOURCES CITED
  • Incident review · 7 Apr — duplicate writes traced to retries
  • ENG-318 — retry limit descoped from spec
  • Agent trace · deploy-bot — flagged as still open
SALES
What did we actually promise Northwind?
WHERE IT LIVES TODAY
CRM notesan email threadthe AE who left
HOVER FOR THE ANSWER
THE ANSWER

A 60-day payment exception, agreed verbally on 27 Feb and confirmed in writing on 3 Mar. It supersedes the standard terms in the original quote.

Quote v1 · 30-day termsFeb
REPLACED BY
Addendum v2 · 60-day3 Mar
3 SOURCES CITED
  • Deal notes · 27 Feb — exception agreed
  • Contract addendum v2 — 60-day terms
  • Superseded: quote v1 — standard 30-day terms
COMPLIANCE
Which policy version approved this claim?
WHERE IT LIVES TODAY
SharePointfour policy PDFsRisk's inbox
HOVER FOR THE ANSWER
THE ANSWER

Policy v2.3, retired on 14 Mar. The exception was signed off by Risk on 2 Mar and replaced by v2.4 on 8 Apr.

Policy v2.3retired 14 Mar
REPLACED BY
Policy v2.4active 8 Apr
3 SOURCES CITED
  • Policy register — v2.3 active 2 Jan to 14 Mar
  • Risk sign-off · 2 Mar — named approver on record
  • Superseded by v2.4 · 8 Apr
SUPPORT
Has this customer been told something different before?
WHERE IT LIVES TODAY
old ticketsa pinned macrotribal memory
HOVER FOR THE ANSWER
THE ANSWER

Yes — twice. The refund window was quoted as 30 days in January and corrected to 14 days in March. The March answer is the current one.

30-day refund windowJan
REPLACED BY
14-day window · v3Mar
3 SOURCES CITED
  • Ticket #4412 · Jan — 30-day window quoted
  • Correction · Mar — policy updated to 14 days
  • Current: refund policy v3
LEADERSHIP
What has changed since the last review?
WHERE IT LIVES TODAY
last board deckthree spreadsheetsFebruary's notes
HOVER FOR THE ANSWER
THE ANSWER

Nine decisions recorded, two of them reversals, and three commitments still open past their date. Both reversals sit in pricing.

9
DECISIONS
2
REVERSALS
3
PAST DUE
3 SOURCES CITED
  • Pricing review · 8 Apr — Mar framework withdrawn
  • 9 decisions recorded since 26 Feb
  • 3 commitments past due
WHERE THE ANSWERS COME FROM

Not your inbox. Memuron answers from what your own agents, tools, and services write into it.

SDKMCPAPIDOCUMENTS

You decide what goes in. Memuron decides what stays true.

WHAT YOU GET

Answers you can trust, knowledge that sticks, and control you can prove.

01

Answers you can trust

Contradictions are settled the moment something is written, so what comes back out is what your organization actually decided.

Finds the right answer, not the closest text

Vector, lexical, and link-description matches surface the memory plus the edge that made it relevant.

Follows the connection, not just the keyword

Ask a question and walk the edges that answer it — not just the text that matches it.

node→ caused_by →nodehops ≤ 2

Hands your model a clean, cited answer

Prompt-ready blocks with citations, graph links, breadcrumbs, and a hard budget.

[M1][L1]
02

Knowledge that sticks

Corrections last. Sources stay attached. What your AI learned last quarter is still there this quarter, instead of resetting every session.

Wrong facts can be fixed, not buried

live · hover a node

Memuron stores every relationship as a first-class memory_link, including self-loops and parallel edges agents can inspect and correct later.

Every fact keeps its receipt

Attach custom_id, session, thread, source ID, and URL to every memory or document.

Point it at your files and links

Fetch pages, sync local folders, and keep files mapped to stable graph nodes.

URLMDPDFDOCXXLSX
03

Control you can show an auditor

Separation between customers, a full history of every change, and rules for what is kept and for how long — so adopting AI does not mean losing oversight.

One customer's data can never reach another's

Cryptographic tenant_id boundaries plus spaces for project, team, and user context.

org:acme
••••
org:globex
••••

Replay any change, any time

Every create, update, and link is a typed, append-only event you can replay.

memory.createdmemory.updated

Keep what you must, delete what you must

Scope what each space can see, and retain or remove memory by scope.

space.workretain · remove
FOR DEVELOPERS

MCP native. Connect Cursor and other agents over an authenticated HTTP stream — no key wrangling. Full SDK and API in the docs.

Read the developer docs
HOW IT WORKS

Five moves, one graph.

Follow a single memory as it moves through Memuron — connect, ingest, verify, recall. Then a fifth move that turns stored context into something you can actually trust.

  1. quickstart
    $ pip install memuron
    SDK · MCP · REST
    01 / CONNECT01

    Plug in, in minutes.

    One SDK, MCP server, or REST call. Runs anywhere your agents run — edge functions to long-lived services.

  2. spec.pdf
    thread
    tool_call
    02 / INGEST02

    Feed it any context.

    Docs, chat threads, tool calls, decisions. Memuron extracts and types the content for retrieval — no schema wrangling.

  3. Guardian · verified
    mergelinkduplicate
    03 / VERIFY03

    The Guardian checks it.

    Before anything lands, the Guardian checks it against what you know: dedupe, merge, link, or reject. Not a blind vector dump.

  4. why did onboarding stall?
    04 / RECALL04

    Query one living graph.

    Memory, RAG, and profiles share a single semantic graph your agent traverses — not three systems you stitch together.

  5. append-only ledger
    agent · claudeconfirmed
    human · reviewcorrected
    source · spec.pdflinked
    05 / COMPOUND05

    Every recall is provable.

    Source, actor, and full history land in an append-only ledger. Corrections compound — the next decision starts from what you've verified.

Steps 3 and 5 are the difference. Anyone can store and retrieve vectors. Verifying what enters and proving what came out is what makes the memory worth building on.

WORKS WITH WHAT YOU HAVE

No new app for anyone to learn.

Memuron sits behind the tools your teams already use. Connect it once, and what one assistant learns is available to the next one — same memory, same rules, no copying context between windows.

PROOF & AUDIT

Prove what your AI knew, and why it acted.

Every decision leaves a trail. Rebuild it from the sources, the corrections, and what actually happened — narrowing with Retrieve, Trace, and Confirm — so risk, compliance, and leadership get a straight answer instead of a guess.

AUDIT QUERYorg.acme/ops.deploys

Which rollback decisions were authorized after recent production incidents?

Retrieve"deploy decisions"Tracecaused_by · 2Confirm"rollback"
EVIDENCE STANDARDS

Before a decision can stand as institutional evidence, it has to clear five bars. Memuron is built so every memory can.

  • Authenticity

    Every memory carries who wrote it, from what source, in which session, and when — so you can prove it is what it claims to be.

  • Integrity

    Nothing is silently overwritten. Creates, corrections, and links are typed append-only events — the record stays tamper-evident.

  • Completeness

    Prior versions, linked sources, and corrections stay in the trail — the full decision history, not a cherry-picked snippet.

  • Reliability

    The Guardian verifies before write; hybrid search and graph recall use governed, inspectable methods — not opaque vector dumps.

  • Strict Chain of Custody

    Who wrote, who verified, what changed, and who recalled — a chronological trail you can replay from first write to last audit.

EVIDENCE PIPELINE·click a stage to narrow the trail
3 decision records· 0.018s

Retrieve records that mean “deploy decisions.”

Meaning-ranked retrieval surfaces institutional knowledge about shipping — even when the wording differs across teams and systems.

6 candidates
ranked by relevance
DECISION LEDGERauto-cycling · hover to inspect
evt_4f2Rollback of v2.3 authorized after error-rate spike0.91
evt_9a1Deploy gate requires successful canary before promotion0.84
evt_c07Incident: production latency regression under review0.79
evt_2b8Rollback policy now requires change-board sign-off0.76
evt_7e3Standup notes — sprint planning0.61
evt_a15Weekly deploy cadence adopted as operating standard0.58
SECURITY & DEPLOYMENTENTERPRISE PLAN

Memuron runs where your data has to stay.

On-prem, in your own cloud, or fully air-gapped — same API, same graph, same append-only ledger. Wherever it runs, the memory and its full audit trail stay inside your boundary.

YOUR PERIMETER · your racks
01 / ON-PREM

In your data center.

Self-host the engine on bare metal or your own Kubernetes. The graph and the ledger never leave your perimeter.

YOUR PERIMETER · your account
02 / YOUR CLOUD

In your VPC.

Bring your own cloud on AWS, GCP, or Azure — provisioned inside your account, under your keys, from day one.

YOUR PERIMETER · no egress
03 / AIR-GAPPED

Fully offline.

Run the whole stack air-gapped with zero outbound calls — for classified, regulated, or sensitivity-first work.

Available on Enterprise plans. Reviewing Memuron for a regulated or security-sensitive deployment? We'll walk your team through isolation, data residency, retention, and self-host options.

Talk to an enterprise architect
THE LIVING LEDGER

Your Agent
remembers what your organization already verified.

Every verified decision compounds into shared, governed memory — queryable across teams, agents, and sessions, never from scratch.

Memories under governance
memories under governance
Contradictions reconciled · 24h
UNDER THE HOOD

No lock-in on the memory your business runs on.

Memuron is the chassis; Arthaanu is the engine underneath it — and it is open source. You can read the code, run it yourself, or take your memory with you. Nothing about the knowledge your company builds up is locked to a vendor, including us.

THE ENGINE BLOCK
ArthaStoreappend-only ledger

the append-only source of truth — objects, encodings, relations and traces

ONE LEDGER · EVERY INDEX REBUILDABLE
semantic_eventsappend-only
SEQEVENTMESSAGE
Vector0
sqlite-vec · cosinerebuilding…
Lexical0
FTS5 · BM25rebuilding…
Graph0
edges · traversalrebuilding…
Temporal0
point-in-timerebuilding…
SIX WAYS TO HOLD A MEANING — THE COMPRESSION GRADIENT
The same memory, re-encoded left to right — each arrow is a recorded transform.
01token
rolledbackv2.3deploy
extract
02tag
#deploy#rollback
link
03relation
caused_by →
assert
04claim

“v2.3 deploy was rolled back.”

embed
05embedding
[ 0.12, -0.04,
0.88, 0.31,
-0.17, … ]
summarise
06summary
tokencompression · very highloses context

See it against your own stack.

Book 30 minutes. We'll map Memuron to your AI stack, your governance rules, and where your data is allowed to live — before you commit to anything.