Six layers.
One memory. Built for sovereign deployments.
Meshline is not a search index, a chatbot, or a wrapper around a foreign API. It is the substrate beneath all three: a permission-aware evidence graph that knows what your organization knows, who is allowed to know it, and what just changed.
Every enterprise generates an unbroken stream of decisions, documents, signals and conversations. None of it is connected. None of it is queryable. None of it is alive.
Your organization knows everything.
It just can't remember.
A 10,000-person enterprise produces, on average, ~6 petabytes of internal knowledge per year, scattered across thirty-plus systems, locked behind incompatible permissions, decaying inside private channels and personal drives.
Search retrieves documents. Document AI summarizes them. Neither knows what the company decided last Tuesday, who contradicted whom in the architecture review, or which strategic initiative quietly drifted off-mission six months ago.
CAPTION: The same fact appears in seven different systems under seven different schemas. A meeting decision contradicts a Slack thread that contradicts a Confluence page. No one notices.
Meshline does not search your data. It rebuilds it, into a permission-aware graph of evidence-grounded claims, decisions, entities and relationships.
The difference is not incremental.
It is structural.
From files → to facts.
Every document, message and decision is parsed into evidence-grounded claims with source-level lineage.
From keyword search → to context.
Queries return synthesized answers with full provenance, not ten ranked links you have to read.
From dead archives → to live memory.
New evidence updates existing claims. Contradictions surface. Stale beliefs are flagged automatically.
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.
Application Surface
A single memory powers every channel. Executives query the company. Employees query their work. Auditors query the trail.
Context Kernel
The kernel mediates every interaction. Agents never see raw memory. They receive permission-filtered, deduplicated context packets compiled to a token budget.
Memory Interface Layer
LOCKEDThe graph substrate and the snapshot system, custom-built for memory workloads. This is the layer that makes the rest possible.
Standardization Pipeline
Documents → layout-grounded extraction. Audio → transcripts + speaker + acoustic events. Video → tracks + scenes + masks. Code → AST + symbol graph. Every observation cites its source coordinates.
Crawler & Ontology Engine
A frontier-class model surveys the corpus, discovers domain concepts, resolves entities, induces relations and proposes the ontology, dynamically, from your data, never templated.
Source Connectors & Evidence Vault
Immutable, hash-addressed originals. Permissions inherited verbatim from source systems. Every byte versioned, every access audited.
Three principles
non-negotiable.
Meshline’s commercial moat is not a model. It is the layer between the model and the data: the ingestion pipeline, the custom graph substrate, the context kernel. We deliver it through direct engagement with each customer.
The proprietary core.
Delivered through engagement.
- L1 · Object storeFiles, blobs, mailflow, recordings
- L2 · LakehouseIceberg / Parquet, columnar truth
- L3 · Search + vectorLexical, semantic, hybrid recall
- L4 · Graph topologyEntities, relations, lineage
- L5 · Context kernelStateful, permission-aware activation
Custom graph substrate
Built for memory workloads, not generic graph queries. Hot activation, snapshot reuse, pointer-based context delivery. Faster than off-the-shelf graph databases because it is not a graph database. It is a memory system that uses graph topology.
Multi-pass crawler
Six deterministic passes plus continuous incremental ingest. Modality decomposition, ontology induction, entity resolution, claim extraction. Driven by a frontier model behind a model-agnostic gateway. Tunable to the customer's domain.
Stateful context kernel
Tracks every active session's pointer set, suppresses duplicate context, compiles delta packets to a token budget, and enforces permission inheritance at the edge of every interaction.
Snapshot system
Pre-computed, queryable memory capsules, by entity, project, person, time window. The kernel can stand up working context for any question in milliseconds, not minutes.
Today, Meshline is delivered through direct engagement with each customer, sized to corpus, integrated to source systems, hardened to deployment posture. The Memory Interface Layer is the persistent core that travels across every engagement.A productized form for mid-market enterprise is on the roadmap. The sovereign tier remains scoped per customer.
Meshline's crawler is a six-pass cold-start that turns raw, disordered enterprise data into a typed, contextual ontology. It runs once on your corpus, then continuously on every new artifact.
A knowledge graph is only as honest
as the model that built it.
Source preservation
Every byte hashed. Every file versioned. Every permission inherited. No intelligence yet, just an immutable, audit-grade record of the company's reality.
Modality decomposition
Documents become layout-grounded blocks. Audio becomes diarized transcripts plus acoustic events. Video becomes tracks, masks, scenes. Code becomes symbol graphs. Reality is preserved in modality-native form.
Ontology discovery
A frontier-class model surveys the corpus and discovers what the company actually is, its products, its people, its risks, its acronyms, its decisions in flight. Tags are not assigned. They are induced.
Entity resolution
Aliases like "Project X", "X Inc.", "the startup", "the new initiative" all resolve into canonical entities, with confidence, aliases, evidence pointers and unresolved alternatives preserved.
Concept graph induction
Relations are proposed, validated, and materialized. The ontology is dynamic but governed: nothing enters the graph without provenance and confidence.
Activation snapshots
Pre-computed memory capsules: by project, person, customer, time window, domain. The kernel can stand up working context for any question in milliseconds.
A serious local model requires serious local infrastructure.
This is the room.
Every source the organization already pays for, reduced to one memory.
Multi-pass ingestion converts files, threads, tickets and recordings into evidence-grounded claims with source-level lineage. Permissions are inherited from the source. The kernel re-derives on every change.
- Outlook
- Gmail
- Proton Mail
- Google Drive
- OneDrive
- Box
- Dropbox
- Slack
- Microsoft Teams
- Discord
- Telegram
- Signal
- Notion
- Confluence
- Google Docs
- Jira
- Linear
- Asana
- GitHub
- GitLab
- Bitbucket
- ChatGPT
- Claude
- Gemini
- Salesforce
- HubSpot
- Google Calendar
- Zoom
- Google Meet
- Grafana
- Datadog
- Decisions
- Entities
- Contradictions
- Lineage
- Risk events
- Citations
One memory, many surfaces. Each is permission-aware, evidence-grounded, and continuously running. Each can be silent (surfacing only what passes the threshold) or active, on demand.
The Company Strategizer.
And five experiences around it.
A silent strategist that runs continuously.
The Strategizer ingests every new artefact the moment it lands and maintains a live, bird's-eye view of priorities, risks, customers, vendors, projects and mission drift. It connects dots no human reviewer has time to connect, and surfaces them only when they pass a confidence and importance threshold the executive office controls.
What changed this week.
A weekly synthesis of new evidence, decisions, and shifts across every source the organization runs on.
Where strategy and execution diverge.
Continuous comparison of stated strategy against operational reality. Drift is named, sourced, and ranked by impact.
Contradictions across teams.
Surfaces the policy on page 14 that contradicts the Slack thread that contradicts the OKR doc, with full lineage.
How was this decision actually made?
Reconstruct any decision from the source artefacts: who proposed, who approved, what was overruled, what changed.
Cited briefings produced from the live corpus.
A weekly executive memo with citations, contradictions, and confidence, generated from the working memory itself.
The same memory, exposed through the channels people already use, with permissions inherited from the originating system, never weakened.
On your hardware. In your facility. Under your law.
Strategy. M&A. Personnel records. Counsel correspondence. Operational telemetry. The data that defines an organization demands explicit, auditable guarantees. Meshline scopes them to three deployment tiers; in the sovereign tier, nothing leaves the perimeter.
The crown jewels get a perimeter,
not a promise.
Everything inside the perimeter
In the sovereign tier, nothing leaves the perimeter: no data egress, no outbound model calls, no silent telemetry. Hardware you own, in facilities you control, on networks you operate.
Permission inheritance
Every derived asset (summary, embedding, graph edge, entity record) inherits the access policy of the source it was extracted from. RBAC is propagated, not bolted on.
Hash-addressed evidence
Every byte of source data is hash-addressed and versioned. Tampering is detectable. Lineage is provable. Every model output can be traced to its evidentiary source coordinates.
Mediated agent access
No agent has direct read access to the graph, the vault or the source systems. Agents request context. The kernel decides what, if anything, to reveal.
Immutable audit log
Every query, every retrieval, every model invocation is logged to an append-only store. Regulator-grade transparency. No retroactive edits possible.
Model-agnostic gateway
The multi-pass crawler reaches frontier-class models through a model-agnostic gateway: a model hosted inside your perimeter, a zero data retention endpoint, or any API you approve. In the sovereign tier, weights, prompts and outputs never leave the deployment.
Three tiers. One governance spec.
Permission inheritance, hash-addressed evidence and the audit ledger ship in every tier. The perimeter is the variable you choose.
Meshline does not replace any system the organization already runs. It sits beneath them, making the entire stack queryable, accountable and alive.
The output is not a chatbot.
It is institutional consciousness.
A bird's-eye view of the company.
A single, queryable picture of every active project, decision, risk and contradiction, synthesized live from every source the organization runs on. The board sees what is happening, not what was last reported.
A silent agent, always running.
A continuous background process correlates new evidence against the existing graph. It surfaces inconsistencies, mission drift, contradictions and stale beliefs the moment they appear, to the people authorized to act on them.
Every channel. One memory.
WhatsApp. Slack. Microsoft Teams. Email. Voice. A custom app. Every interface routes to the same kernel, the same governance, the same evidence-grounded answer. No channel sees more than its user is permitted.
A chain of custody on every answer.
Every model output is a traceable claim. Click a sentence, see the source paragraph, the timestamp, the author, the version, the chain of derivation. Auditable by counsel, regulators and the security team alike.
Architectural reasoning, not vendor claims.
Performance numbers vary by corpus, by hardware, by ontology. These four do not. They are architectural constants, fixed at design time, by construction.
The numbers that don't move
across deployments.
In sovereign deployments, no model API calls leave the perimeter: no silent telemetry, no vendor-side weights, and the frontier model that runs the crawl is hosted locally. ZDR and Meshline cloud tiers are available.
Every claim is grounded to source coordinates, page, paragraph, frame, timestamp. Claims without lineage are not produced.
Deployment-scale graph topology. Sized to the customer's corpus, the customer's ontology, the customer's pace of change.
A derived summary, embedding, graph edge or entity record can never be more permissive than the source it was extracted from.
Pure RAG, pure vector, pure graph, pure cache: each fails under sovereign-enterprise load. Meshline is the layered composition that survives the constraints all of them violate.
No single component is sufficient.
The composition is the answer.
· Meshline does not replace any of these components. It composes them through a proprietary kernel that decides what to activate, what to cite, and what to suppress. The graph is one layer. The vector store is one layer. The lakehouse is one layer. The kernel is the answer.
Below ~10,000 entries, every retrieval strategy looks fine. The curves diverge as the corpus grows. This is the single chart that explains why naive RAG fails at sovereign-enterprise scale, and why a kernel that emits deltas, rather than retrieved sets, is the only architecture that holds.
The same query, the same answer.
A different cost curve.
Retrieval-augmented generation pulls text fragments. Meshline reconstructs the underlying object: the decision, the entity, the contradiction, the lineage. The difference shows up the moment a question matters.
RAG was built for documents.
Enterprises run on decisions.
Text chunks
Lost: chunks are independent
Bolt-on filters, often leaky
Approximate: chunk ID
Returns both, ranked by similarity
Treats stale and current as equal
Balloons exponentially
You audit the prompt
Hallucinates plausibly
APPENDIX A · THE GRAPH-DB LANDSCAPESeven engines. Different jobs. Meshline does not bind to one.
Sovereign enterprise memory has three workloads: hot activation under millisecond latency, distributed analytics over the full graph, and embedded snapshots at the edge. Each calls for a different engine. The Memory Interface Layer is what abstracts the choice.
Seven engines. Different jobs.
Meshline does not bind to one.
Neo4j
BASELINE / DEBUG SURFACEMature ecosystem, Cypher, strongest tooling, best visualizations.
Not the fastest at distributed scale; vector + full-text are layered on rather than first-class.
Memgraph
PRIMARY HOT-ACTIVATION GRAPHIn-memory C/C++ engine, Cypher-compatible, built-in vector and text indexes, fast multi-hop traversal.
Younger ecosystem than Neo4j; community-edition licensing requires review for redistribution.
FalkorDB
HIGH-THROUGHPUT ACTIVATIONGraphBLAS sparse-matrix backend; the RedisGraph successor. Exceptional read-heavy traversal latency.
SSPL license requires legal review for OEM deployments; ecosystem is narrower.
TigerGraph
MASSIVE DISTRIBUTED GRAPHAudited LDBC SNB BI results at scale; serious distributed graph analytics.
Operationally heavier; overkill for single-tenant memory at most enterprise sizes.
GraphScope (Apache)
CLUSTER ANALYTICSApache-2.0; audited LDBC interactive results; designed for cluster-scale graph analytics.
Less ergonomic for hot serving than Memgraph or FalkorDB.
Kuzu
EMBEDDED SNAPSHOTSEmbedded architecture, columnar storage, interesting performance characteristics.
Original repository archived October 2025; production use depends on fork ecosystem.
Amazon Neptune
CLOUD-PREM ONLYManaged AWS, supports Gremlin and SPARQL, integrates with the AWS data plane.
Full-text search depends on OpenSearch integration. Disqualified for air-gapped sovereign deployments by virtue of being managed cloud.
One memory. Many engines underneath.
Different deployments choose different engines based on the customer\'s posture, sovereignty, scale, ecosystem, license. The Memory Interface Layer presents a stable graph contract; what runs underneath is a deployment decision, made jointly with the customer's security and platform teams.
APPENDIX B · INSTRUMENTATIONThe numbers we measure. Sized to your corpus.
Performance numbers are deployment-specific. The structure of what we measure is constant across every install. Numbers are produced from the customer's own corpus during phase 0 and travel with the methodology brief that explains them.
The numbers we measure.
Sized to your corpus.
- Cold-start ingest convergenceTime from connector enable until the first answerable query.
- Snapshot-hit context build (p50)Median context-packet latency when the active snapshot covers the query.
- Cold-pass context build (p95)95th-percentile latency when the kernel has to build context from cold storage.
- RAG vector-search baseline (p95)Side-by-side latency for a vanilla vector retrieve over the same corpus.
- Token cost @ 100k entries (delta)Per-query LLM cost compared to a vector + re-rank baseline at 100k entries.
| SCENARIO | MESHLINE | RAG + POST-HOC FILTER |
|---|---|---|
| Source file moved between folders | preserves source ACL on derived assets | leaks via the embedded vector |
| RBAC group renamed at source | propagates within the next ingest cycle | requires re-embedding the corpus |
| Source permission revoked retroactively | derived assets become inaccessible immediately | derived chunks remain queryable |
| Cross-tenant query attempt (multi-tenant deploy) | blocked at the kernel boundary | depends on filter correctness |
| Audit reconstruction over 12 months | lineage chain is queryable end-to-end | no lineage by design |
Numbers come from your corpus, with the methodology that explains them.
We do not publish portable benchmarks. Each customer receives the full p50/p95/p99 distribution measured on their corpus, the configuration that produced those numbers, and the explicit bounds within which Meshline will, and will not, commit to a number.
When the data is sovereign, the machine has to be sovereign too. The Altar is the physical embodiment of Meshline: a purpose-built room that houses the company's living memory.
A room you can walk into.
A memory that never leaves it.
CAPTION: The obelisk is the visible control plane: live ingestion, memory activation, audit telemetry. The compute and storage substrate sits adjacent, behind glass, in the Vault.
Pilot Altar
Proves the Memory OS, ingestion, RBAC and exec UX. Hardware sized for one department, one corpus. Suitable for a 12–16 week prototype.
Sovereign Intelligence Chamber
Full obelisk centerpiece, command wall, separated Vault. HGX B200 / H200-class compute. The visual flagship for an executive deployment.
AI Factory
Rack-scale liquid cooling. Multi-rack fabric. Full disaster-recovery topology. Sovereign inference factory plus full Memory OS at organizational scale.