The Knowledge-Base OS

The memory layerbetween AIand everything you know.

AI models know the world’s public knowledge. Meshline gives them yours: every document, message, decision and signal you produce, turned into one governed, permission-aware memory that surfaces what you already knew but never connected. Run it on your own hardware, through a zero data retention API, or in the cloud.One memory. One source of truth. Cited, auditable, alive.

FIG. 01 · ENTITY RESOLUTION
MEETING NOTE · INTERNALWED 14:32"The company's runway is tightening,retention is weak, and the team isconcerned about product–market fit."The companyCLAIM · CITEDrunway is tighteningCLAIM · CITEDretention is weakCLAIM · CITEDconcerned about product–market fitRESOLVEDSTARTUP X14 CITATIONSVECTOR SEARCHNO MATCH
SOURCE SYSTEMS
  • Outlook
  • Gmail
  • Proton Mail
  • Google Drive
  • OneDrive
  • Box
  • Dropbox
  • Slack
  • Microsoft Teams
  • Discord
  • Telegram
  • WhatsApp
  • Signal
  • Notion
  • Confluence
  • Google Docs
  • Jira
  • Linear
  • Asana
  • GitHub
  • GitLab
  • Bitbucket
  • ChatGPT
  • Claude
  • Gemini
  • Salesforce
  • HubSpot
  • Google Calendar
  • Zoom
  • Google Meet
  • Grafana
  • Datadog
+ recordings · CCTV · IoT · datastores · APIs
A LAYER, NOT AN APP

Meshline is the substrate beneath search, agents and chat: a permission-aware graph that knows what your organization knows, who is allowed to know it, and what just changed.

It cites every claim it produces. It inherits permissions from the source systems it ingests, and it re-derives the truth on every change. In sovereign deployments it runs on your hardware and nothing leaves the perimeter; ZDR and Meshline cloud tiers are available.

DEPLOYMENT TIERS
SOVEREIGN ON-PREM

Everything inside the perimeter. Zero egress.

ZDR API

Inference through a zero data retention API.

MESHLINE CLOUD

Managed by Meshline, governed by your policies.

DESIGN-TIME INVARIANTS
  1. 0
    data egress · sovereign tier

    In sovereign deployments, nothing crosses the perimeter: no model calls, no telemetry.

  2. 100%
    cited output

    Every claim ties back to a source coordinate.

  3. 10⁹⁺
    graph edges

    Deployment-scale topology, sized to the corpus.

  4. 1:1
    permission inheritance

    Derived assets cannot be more permissive than source.

Similarity proposes. Evidence decides.

WORKED EXAMPLE· A SINGLE MEETING NOTE

A short, concrete walkthrough: the kind of artefact every enterprise generates a hundred times a week. The difference is what the system does with it.

What naive retrieval does.
What Meshline does.

MEETING NOTE · INTERNAL · WED 14:32

"The company's runway is tightening, retention is weak, and the team is concerned about product–market fit."

NAIVE RAGCHUNK · EMBED · RETRIEVE
  • Splits the meeting note into chunks
  • Embeds the chunks
  • On query, retrieves the highest-similarity fragment
  • Returns it without context, without lineage, without permissions inferred
What it cannot do: reason that the note is about Startup X, that Startup X is a portfolio company, that runway / retention / PMF are metrics, or that the user asking may not be cleared to see any of it.
MESHLINERESOLVE · LINK · ENFORCE
Resolves the entity
"the company" → portfolio company Startup X (canonical, with confidence and aliases)
Classifies the topic
Startup health: runway, retention, product–market fit
Cross-links the evidence
Last quarter's board minutes, the fundraising memo, the unanswered email from the CFO
Detects the contradiction
A Q1 commitment from Operations that conflicts with what was just said
Enforces the permission
None of this is visible to the user asking unless they have clearance for portfolio-company financials
Result: the agent receives a permission-filtered, cited, evidence-grounded context packet, not a chunk.
03 · ARCHITECTURE· THE STACK

Every layer is auditable. Every layer is governable. The Memory Interface Layer is the proprietary core, delivered through direct engagement with each customer.

Six layers.
One that never has to leave your premises.
One that we built ourselves.

FIG.02 · MESHLINE REFERENCE STACKPER-CLIENT DEPLOYMENT
L06

Application Surface

WhatsApp · Slack · Email · Teams · Web · Voice · Custom Agents

A single memory powers every channel. Executives query the company. Employees query their work. Auditors query the trail.

L05

Context Kernel

Stateful · Permission-aware · Delta-injected

The kernel mediates every interaction. Agents never see raw memory. They receive permission-filtered, deduplicated context packets compiled to a token budget.

L04

Memory Interface Layer

LOCKED
PROPRIETARY · LICENSED PER DEPLOYMENT

The graph substrate and the snapshot system, custom-built for memory workloads. This is the layer that makes the rest possible.

L03

Standardization Pipeline

Modality-native · Provenance-preserving

Documents → layout-grounded extraction. Audio → transcripts + speaker + acoustic events. Video → tracks + scenes + masks. Code → AST + symbol graph. Every observation cites its source coordinates.

L02

Crawler & Ontology Engine

Frontier model · model-agnostic gateway

A frontier-class model surveys the corpus, discovers domain concepts, resolves entities, induces relations and proposes the ontology, dynamically, from your data, never templated.

L01

Source Connectors & Evidence Vault

Confluence · Notion · Slack · Drive · CCTV · IoT

Immutable, hash-addressed originals. Permissions inherited verbatim from source systems. Every byte versioned, every access audited.

FOUNDATION · SOVEREIGN TIERAir-gapped Linux · Sovereign GPU substrate · Hardware HSM-backed RBAC

Three principles
non-negotiable.

PRINCIPLE · 01
Evidence over embeddings.
Every claim Meshline produces is grounded to source coordinates, page, paragraph, frame, timestamp. Embeddings retrieve candidates. Coordinates prove truth.
PRINCIPLE · 02
Permissions inherit, always.
A derived summary, tag or graph edge cannot be more permissive than the source it was extracted from. RBAC propagates through every transformation, automatically.
PRINCIPLE · 03
The agent never sees memory.
Agents request context from the kernel. The kernel decides what to surface. No agent ever has direct read access to the graph, the vault or the source systems.

Agents request context. The kernel decides what, if anything, to reveal.

ARCHITECTURE

A graph, not a search index.

Meshline reconstructs the underlying object (the decision, the entity, the contradiction, the lineage) and exposes it through a permission-aware kernel. Built for sovereign deployments where RAG cannot go.

See how it works
SECURITY & SOVEREIGNTY

Governed by design.

Clearance enforced in the query itself, source bytes hash-addressed and never destroyed, and a sovereign tier where nothing leaves your perimeter. See the redaction in the clearance lens.

Open the security model
EVIDENCE

Three live deployments.

Three sectors. Redacted. Each one surfaced something the organization had already silently accepted. Read the case files.

Read the cases
By invitation

Engagements begin with a confidential briefing.
In person, where possible.

We work with governments and global enterprises by invitation. Each engagement is scoped to the customer. Qualified inquiries receive a confidential response within 48 hours.