Medusa | Anomaly

Anomaly · Context governance

Medusa.

The governance layer between your data and any model.

AI systems are given access to data and memory with no record of what they saw, or why. Medusa decides what context may reach a model, for which purpose, and keeps a replayable receipt of every decision.

  • Medusa kernel in the application
  • Medusa Gateway between systems and models
  • Medusa Console one control surface
Medusa: a translucent blue jellyfish
Distributed, not central

Enforcement happens wherever context leaves. Policy and receipts are shared across every point, so a purpose denied once is denied everywhere.

Nothing is replaced

Medusa attaches to what already exists: MCP servers, internal APIs, memory stores and on-device data.

Accountable by design

Every delivery can be replayed: what was used, what was withheld, and which rule decided.

Better answers

The model sees the current revision, only what the purpose needs, and memory a person has approved. Smaller packs also mean faster, cheaper answers.

Where it fits

One layer between evidence and inference.

Sources stay where they are. Models stay interchangeable. Medusa decides what context may leave, and for what, and records the decision. The Console governs every point of enforcement from one place.

Medusa architecture Five sources on the left connect to the Medusa kernel in the middle, which applies policy, builds a context pack, records a receipt and explains it, then delivers to two models and an agent on the right. The Console sits above and governs everything. Medusa Console Connections · Policies · Receipts · Explain · Approvals · Keys SOURCES MEDUSA MODELS & AGENTS Document storepolicies, SOPs · read Internal APIsHR, CRM via OpenAPI · read MCP serverstools and resources · govern Memory storeMem0, pgvector · operate On-device datavia the kernel, in process Medusa kernel 1 · Policy decision 2 · Context pack 3 · Receipt 4 · Explain allow · deny · redact · escalate under a named policy version Model Ahosted API Model Bprivate deployment Agentacts with a pack, not access models propose changes · people approve
Embedded

Medusa kernel

Built into the application: Swift for Apple platforms, Kotlin for Android, Rust elsewhere. It works without a network or a model.

  • Governs memory where it lives, including on the device
  • An Explain API and a ready-made "Why this?" view
  • In use inside Moonlight
readgovernoperate
In the path

Medusa Gateway

Attaches to existing systems and enforces at the point where context leaves for a model.

  • MCP first: governs tool results and resources item by item
  • APIs found through OpenAPI, approved by a person
  • Existing memory stores gain provenance, approvals and verifiable forgetting
readgovernoperate
Control

Medusa Console

One control surface over every kernel and gateway. This is what a team operates.

  • Policies with simulation against past requests
  • Receipts search, replay and explain
  • Approvals queue for changes to memory
configauditapprove

One delivery

What happens when an assistant asks for context.

An HR assistant answers an employee's question. Step through what Medusa does before any model sees it.

Employee asksHow much parental leave do I get?

Context request

principal
svc-hr-assistant
on behalf of
employee E-2041
purpose
policy-lookup
budget
2,000 tokens
policy
hr-policy.v7
model
not yet chosen
Context packctx_9f2c…a71e
3 items · 1 field redactedexpires 18:00

Candidate evidence

Parental leave, section 4.2Document store · HR Policy rev 7pendingCurrent revision, matches purpose.
Parental leave, section 4.2Document store · HR Policy rev 6pendingSuperseded by rev 7 on 1 Sep.
Start date and work locationHR API · employee E-2041pendingNeeded to calculate eligibility.
Salary bandHR API · employee E-2041pendingSame record, field not needed for this purpose.
Manager notesCRM over MCP · employee E-2041pendingPurpose mismatch: notes are scoped to performance-review.
Prefers short answers with a sourceMemory store · approved 3 OctpendingApproved preference, allowed for any HR purpose.

Accountability

The same receipt, for a reviewer and for a person.

Reviewers see every delivery in full in the Console. The person an answer concerns sees the same record in plain language, with the means to correct it.

Console › Receipts › rcp_20261120_0417

Parental leave answer

20 Nov 2026 · 14:12
Purposepolicy-lookup
Policyhr-policy.v7
Delivered toModel B
EvidenceSource revisionDecisionRule
Parental leave, 4.2hr-policy@rev7 · 3be1…allowpurpose match
Parental leave, 4.2hr-policy@rev6 · 91d0…denysuperseded
Start date, locationhr-api/E-2041 · 77ac…allowfield scope
Salary bandhr-api/E-2041 · 77ac…redactfield scope
Manager notescrm-mcp/E-2041 · c402…denypurpose mismatch
Answer preferencememory/pref-118 · 0e5f…allowapproved memory
Replay this deliverySimulate with hr-policy.v8Export receipt
9:41●●●
Play something for deep work
Starting "Low Tide" from your Focus mix. Instrumental, about 60 minutes.
Why this?

Why this?

Used
  • Your saved Focus preference, 12 Nov
  • Your last 3 focus sessions on this Mac
Not used
  • Messages: not allowed for music
  • Location: not needed
CorrectForget

Governed and explained on the device. Nothing in this explanation left it.

Position

Gateways govern calls. Memory layers store. Medusa governs what reaches the model.

CategoryExamplesGovernsMisses
MCP and AI gatewaysTrueFoundry, Kong, Traefik Hub, DockerWhich agent called which tool, allowed or deniedWhat content reached the model, and why
Memory layersMem0, Zep, LettaStoring and retrieving memoryProvenance, purpose limits, approvals, proof of forgetting
ExplainabilityMostly researchModel internalsReliable tools for explaining generations
MedusaKernel, Gateway, ConsoleEach item of context: source revision, purpose, policy, what was withheldDoes not explain model weights. Explains model inputs, the same way every time.

Availability

An open contract, and a platform to run it.

The contract is open. Organizations pay for what makes it hold in practice: receipts, explanation, the Console, support, and operation inside their own boundary.

OpenACRP schemas

The contract Medusa implements. Versioned JSON schemas and a spec anyone can adopt.

FreeApache-2.0, for anyone
DevelopersMedusa kernel

SwiftPM, Maven, crates.io. Runs on the device, with no metering.

Freeunder $5M annual revenue, then $3,000 per app a year
TeamsMedusa Cloud

Receipts hold references, not content, so hosting the Console adds no exposure.

$399 a month250,000 governed deliveries, then $12 per 100,000
OrganizationsMedusa Enterprise

The full platform, self-hosted, with customer-held keys, kernel licences and support.

$40,000 a yearunlimited deliveries
  1. Receipt ledger and replay. Run the request, pack, decision and receipt contracts in the kernel.
  2. Explain API and "Why this?". Swift first, with Moonlight as the reference application, then Android.
  3. Attach layer. MCP proxy, then APIs with discovery and approval, then memory-store overlays.
  4. Console v1. Connections, Receipts and Explain; then Policies with simulation and Approvals.
  5. Hardening. Authenticated writes, encrypted store, indexed retrieval, policy-engine adapter, signed releases.
  6. Reference deployments. Moonlight on the kernel, Skygate behind the Gateway, Paradox as a memory overlay.

Building with AI memory or agents?

Tell us what your boundary has to satisfy. Early access starts with a conversation.