Constellation
Agent mesh & governed context
The agent mesh and shared org memory, spoken over MCP.
A centralized context and knowledge plane. Agents and engineers check in over MCP and recall org facts and ingested docs, scoped by division and role resolved from your identity provider (IdP), enforced twice from one source of truth.
- MCP context server: curated Memory + ingested DocumentChunk corpora, one ACL.
- IdP-group-scoped access; multi-tenant with tenantId as the outermost gate.
- Prompt-injection content-safety layer on every write and recall.
Shared agent memory is a data-leak path with a friendly name.
The moment agents share institutional context, you inherit an access-control problem: a CTO's agent should read engineering and exec knowledge but never HR; a contractor's agent should see less. Bolt memory onto agents without enforcing who-can-see-what, and you have built an exfiltration path and called it a feature.
The questions your reviewers will ask.
“Who is allowed to recall what?”
Access is scoped by division and role resolved from your IdP groups, on both reads and writes.
“Can the search leak what the ACL forbids?”
The same ACL gates CRUD and the vector search, cross-validated so the two can’t drift apart.
“Can one tenant reach another’s data?”
Multi-tenant with the tenant as the outermost gate: no admin escape hatch, IDOR-proven.
“Can a poisoned document hijack an agent?”
A prompt-injection content-safety layer runs on every write and every recall.
“Where do org rules live?”
A governed rules store agents query for policy and conventions, under the same ACL.
Evidence, not assurances.
What a security leader walks away able to demonstrate to a board or an auditor.
- A memory layer where every recall is authorized against the identity your IdP already trusts.
- One source of truth for access, enforced twice and cross-checked.
- Prompt-injection defense on the content agents actually read.
- Protocol
- MCP
- Access
- IdP-group-scoped (division + role)
- Tenancy
- Multi-tenant, tenant-outermost
- Stores
- Memory + DocumentChunk, one ACL
- Safety
- Prompt-injection content layer
One module of the Alyria platform.
Constellation works alongside the rest of Alyria — prevention and detection for every AI agent your people run, tied together by one policy engine and one signed audit chain.
Put Constellation to work.
Deploy Beacon read-only and see how Constellation fits at the endpoint, under your keys.