Documentation
Verity developer docs
Verity is an open-source, permission-aware shared context plane for enterprise AI agents. It mirrors your systems of record into a bi-temporal memory store, inherits their ACLs into a permission graph, and serves every agent scoped hybrid recall in under 50 ms — with scoping enforced in the index, never delegated to the model. It is MCP-first and Apache-2.0.
Start here
Boot Verity and run a leak-proof query in five minutes
One opinionated, copy-paste path from an empty laptop to a permission-filtered query — ingest a real directory, watch a second agent get zero results on the same data, and wire an agent over MCP. Every decision is already made for you.
Get started → New to Verity? Take this first. It makes you successful before you ever read a reference page.What Verity is
Enterprises run agents across sales, support, marketing, and ops, but each agent is an island: the context it needs lives in CRMs, ticketing systems, docs, and wikis. Verity is the shared memory that connects them — and the reason it is safe to share is that every retrieval is pre-filtered by the caller's permissions inside the index query itself. Three measured claims hold the product together (the honesty log):
- Provable scoping. An agent talking to customer A can never surface customer B's pricing — enforced in-index, fail-closed. 0 leaked items across 1220 adversarial probes.
- Live truth. A source change is queryable in milliseconds — a deterministic keyed upsert, no LLM in the write path. Measured CDC-to-queryable: 31 ms.
- Inner-loop speed. Every retrieval path stays inside <50 ms p95 at 1,000,000 chunks; point reads are ~0.5 ms.
The mental model behind it is four ideas — layered bi-temporal memory, entities and tags, scopes and handles, and the cross-customer knowledge layer. Read the full concepts page →
Find your page
The rest of the docs are organized by what you're trying to do. Start at the top; drop into Build when you're wiring real systems, Understand when you need the guarantees.
When you need Verity
Seven concrete failure modes of RAG-over-everything and hand-rolled filters — and an honest test for where you do not need Verity at all. Read this if you're evaluating.
who evaluators and skeptics. when you're deciding whether this class of problem is yours. 5 minGet started
The prescriptive tutorial: boot, ingest, query, watch the scoping wall hold, connect an agent. Copy-paste, every decision made for you.
who everyone, first. when you want a working, scoped memory before reading anything else.Connecting your systems
Per-system setup, first data in — HubSpot, Salesforce, Google Drive, Postgres/MySQL, a doc corpus, any internal system.
when you have a real source (“I have a HubSpot — walk me through it”). deep diveIngestion
The BYOT doctrine and every entry point — CLI, envelopes, webhooks, files, framework sinks, MCP tools, connectors, manifests.
when you need the mechanism reference for a specific door into the write path. recipesCookbook
Copy-paste recipes, one goal each: scope an agent to one account, ingest a HubSpot account, connect an agent over MCP, and more.
when you know what you want and just need the exact commands. referenceAPI reference
Every REST route grouped by concern, and all 14 MCP tools with arguments, example requests, and responses.
when you're coding against Verity and need exact routes and parameters.Concepts
The mental model with diagrams: bi-temporal L0/L1/L2/L3 memory, entities and tags, scopes and handles, the knowledge layer.
when you want the whole picture before the details. the thesisPermissions
Scoped in the index, not the prompt. Principals, entity scope, confidentiality, HMAC handles, fail-closed semantics, scoping recipes.
when you need to say precisely who can see what, and prove it. differentiatorKnowledge layer
Learns across your customers, never leaks them. De-identification gates, the §7g carve-out, the merge cascade with measured precision/recall.
when you want org-wide learning, not just per-customer recall. workedUse cases
End-to-end scenarios: a leak-proof support agent, two agents sharing one account, org-wide learning, a 60-second webhook source.
when you learn best from a full, concrete example.Load-bearing principles
These recur throughout the docs; internalize them once.
- Scoped in the index, not in the prompt. Out-of-scope memory never reaches the model. Prompt injection cannot widen visibility because query text is just data.
- Fail closed, everywhere. Empty principal set → nothing visible. Unmappable ACL → quarantine, never permissive indexing. Missing scope → empty result.
- Visibility is never guessed. Every entry point demands an explicit visibility decision; its absence is a refusal or a quarantine, never an assumption.
- Every fact carries its ACL provenance. mirrored approximated admin-assigned quarantined — so “how much do I trust this filter” is visible on every hit.
- Measured, never quoted. Latency and leakage numbers come from
verity-benchat a stated corpus size and machine, on an append-only log.