Agent runtimes
AWS Bedrock AgentCore, LangGraph, Temporal
- Entity-generic
- Bitemporal
- Per-turn brief
- In-tenant
- Live-eval
- BYO agent
We're the shared memory and identity layer that sits under a company's AI agents — so every agent knows who's who, remembers what happened, and can prove where each answer came from, all inside the company's own systems.
The problem
Agents run on scattered, stateless, unsourced data.
The engine
One layer resolves identity, remembers with provenance, and governs — underneath any agent.
The edge
The only platform native across all six runtime capabilities.
Traction
Signed design partners in regulated verticals; building toward production.
The ask
Raising a pre-seed to convert the signed pipeline.
MiddyMind is the proprietary entity-resolution and memory engine AI agents are built on and run on. Every entity across your systems, resolved. Remembered with provenance. Served to any agent, per turn, inside your own tenant.
Customers bring their own agents - authored in plain language or as their own code. MiddyMind supplies resolution, bitemporal memory, runtime governance, and a live improvement loop the agent consumes but never has to build.
TL;DR
The entity-resolution and memory engine under any agent — governed, bitemporal, in-tenant. In stealth, building across 2 regulated verticals. Raising pre-seed.
Companies are deploying agents on top of scattered, stateless, unsourced data - and getting confident, ungoverned, unprovable answers at scale.
The failure is structural and lives below the model. It recurs for every agent a company ships.
Signal about any entity is scattered across many systems; an agent sees slices.
Agents are amnesic across sessions and channels; what one learned or promised, the next never sees.
Answers carry no source; when one is wrong, no one can say which system it came from or when it was last true.
Residency, consent, retention, audit - renegotiated per source and per project instead of enforced once, underneath.
Five primitives. Each a live engine module. The agent is pluggable and lives above this line; everything that makes it useful lives below.
identity/
Resolves any entity type across sources.
memory/
Bitemporal (valid-time + transaction-time) store of typed, sourced facts. Now legible and self-improving: correctable Memory Ledger, nightly Consolidation, orthogonal scoping, standardized recall benchmark.
conductor/signal_brief.py
A compact ~200-token typed snapshot.
governance/
In-tenant federation, per-fact attribution, per-purpose consent, three autonomy tiers. Hard cost caps with model-independent, cost-tiered routing and a budget breaker.
eval/, learning/
“The benchmark is yesterday.” Nightly sampling of real interactions.
The spine that ties the five primitives together. Every inbound message flows through one ordered lifecycle. Swap the model and the resolved entity graph compounds underneath.
Two hooks for the agent - read a Signal Brief on demand, write its outcome after. No plumbing. Omnigent swaps Claude / OpenAI / Gemini behind an identical brief.
Two verticals in active build — both in healthcare — with a third planned to prove the substrate is entity-generic. Each brought its own agents and domain; MiddyMind supplied everything underneath.
Medical-device safety intelligence
Design partner · customers signed & in queue
Entity: device
Fuses MAUDE reports, recalls, complaints, service logs, and EHR adverse events into one device Safety Graph — catching harm signals before they cascade.
You'll see: 3 standing agents — procurement risk scoring, complaint intake, and patient-facing safety reports — with automated MDR/QMSR intake and every claim sourced to the exact MAUDE / recall fact.
US-based global health safety intelligence company.
Traction: Paying customers today, more signed and in queue.
Compliance-grade artifact generation
In talks with customers
Entity: application (composite)
Turns an application's architecture, data flows, and controls into signable, regulator-grade compliance artifacts — every paragraph cited and confidence-labeled.
You'll see: 19 compliance tools · 10 seeded agents (orchestrator + 7 specialists + cloud pair) · 5 regulatory regimes (FDA SaMD · HIPAA · IRB · ONC HTI-1 · state AI law). Reg-diff queues only rule-changed sections; CloudOps adds cloud posture (CSPM) and governed JIT access as sourced compliance signals. Ungrounded claims withheld, not guessed.
Two authoring modes, one engine binding. Both select an entity type and connectors, set policy (autonomy tier + cost cap + consent purpose), and deploy onto the harness in-tenant.
A natural-language description is turned into a readable AgentClient subclass by a.
Claude Agent SDK, LangGraph, a container image, or a custom HTTP endpoint.
In stealth. Verified in code, nothing yet in production. Raising a pre-seed round and applying to Microsoft for Startups to underwrite the next twelve months of in-tenant infrastructure.
For VC reviewers: a defensible substrate — entity-genericity, correctable Memory Ledger, closed eval loop — under whichever agent runtime wins.
For cloud-fund reviewers: in-tenant deployment across three sovereign channels, with cost-tiered routing and a per-tenant budget breaker governing consumption.
Fixed 500-item synthetic corpus · run nightly against Memory Ledger head · reproducible eval.
Assumptions & definitions
Stealth-appropriate signals · no revenue disclosed · no logo use without partner sign-off · full definitions & as-of dates in § 12.
Six runtime capabilities, six categories
Every agent platform must own the six columns end-to-end. We scored the closest neighbors in each category against the same runtime bar we hold ourselves to.
“Coverage” counts only Native marks — Partial means the piece exists but ships as a library, service boundary, or manual step rather than a governed primitive. Hover any cell for the definition.
Closest neighbor
Zep / Graphiti covers 4 of 6 capability dimensions. Strongest single competitor, still not a runtime: no entity-generic resolution, no closed eval loop, no bring-your-own-agent harness. No rival — including the doc-AI acqui-hires, the RPA incumbents, and the cloud-ops / CNAPP tools (Wiz, Orca) — is native across all six.
| Category | Entity-generic | Bitemporal | Per-turn brief | In-tenant | Live-eval | BYO agent |
|---|---|---|---|---|---|---|
Agent runtimes AWS Bedrock AgentCore, LangGraph, Temporal | ||||||
Memory layers Zep / Graphiti, Letta, Mem0, Cognee | ||||||
Identity / CDP Amperity, Salesforce Data Cloud | ||||||
CX agents Sierra, Decagon, Intercom Fin | ||||||
Enterprise doc-AI / RAG Glean, Hebbia, Clarative (→ OpenAI) | ||||||
Agentic RPA / orchestration Automation Anywhere, UiPath, Blue Prism | ||||||
MiddyMind This engine |
AWS Bedrock AgentCore, LangGraph, Temporal
Zep / Graphiti, Letta, Mem0, Cognee
Amperity, Salesforce Data Cloud
Sierra, Decagon, Intercom Fin
Glean, Hebbia, Clarative (→ OpenAI)
Automation Anywhere, UiPath, Blue Prism
This engine
Sized top-down from analyst forecasts, bottom-up from the buying centers we already sell into. Numbers are dated, sourced, and footnoted — no round-number storytelling.
The thesis: agent spend is real and about to overshoot; a governed memory + orchestration substrate is what turns the 40% cancellation rate into deployed revenue.
Rules we followed: primary analyst releases only; every figure dated and geo-scoped; forecast years quoted as published (no interpolation); no vendor-sponsored TAMs; MiddyMind revenue projections are not derived by applying a capture rate to these pools — they're built bottom-up from signed design-partner ACVs and disclosed separately.
Four things competitors cannot copy
MiddyMind is the entity-resolution and memory layer under the agent — complementary to the runtime, identity, and memory tools you already run. Read for two reviewers: VC partners underwriting a defensible substrate, and cloud fund leads underwriting governed in-tenant consumption.
One resolution engine for customers, devices, applications, ventures — a surface neither agent runtimes nor CNAPP/cloud-ops incumbents (Wiz, Orca) touch.
The runtime enforces residency, consent and audit inside the customer's perimeter — the substrate cloud programs can underwrite as sovereign consumption.
Bitemporal ledger with nightly consolidation, orthogonal scoping, and a standardized recall benchmark — graded corrections feed a federated reward model.
Model-independent routing with hard per-tenant cost caps — the governance surface cloud funds care about, and a moat no memory or RPA vendor ships.
Deployment postures + pricing shape
MiddyMind is a PaaS you run inside your tenant. Pick a deployment posture below; the pricing shape follows.
Tenant isolation via Postgres row-level security on a per-request tenant GUC. The runtime federation guard refuses to start if the DSN looks like a vendor-owned shared cluster.
Startups, non-regulated pilots
Our cloud.
Default · regulated enterprise
Runtime federation guard refuses to start against a vendor-owned shared cluster.
Government · defense
Offline install, signed traces only for remote diagnostics.
Economic buyer
Customers bring their own agents; MiddyMind is sold embedded under partners like EmpowerReg and Cleared AI, who carry it into the regulated buyer. The partner signs the commercial contract; MiddyMind is the substrate underneath.
EmpowerReg
At a Class II/III medical device maker — owns MDR/QMSR obligations and signs for the safety-intelligence stack that EmpowerReg delivers.
Cleared AI
Shipping a regulated software application — owns the artifacts a regulator or enterprise buyer will sign, and signs for the assurance package that produces them.
MiddyMind
Memory + governance substrate
EmpowerReg
Regulated SaaS partner
End customer
Device maker / hospital / regulator
MiddyMind runs underneath; the partner owns the customer relationship.
Pricing shape
You rent the engine, we run it inside your tenant. Two co-designed usage counters are metered in the runtime today.
Early design partners get the platform floor waived in exchange for a revenue-share agreement on downstream revenue their MiddyMind-powered agents generate. Both sides win as the substrate compounds.
| What you pay for | Design partnerNow | Standard PaaSAt GA |
|---|---|---|
| Platform floor | Waived. No annual fee during the design-partner phase. | Annual fee by deployment posture — shared, in-VPC, or air-gapped. |
| Usage — resolved entities | Not metered. You keep everything you build. | Metered per entity under management. |
| Usage — agent-turns | Not metered. | Metered per turn served. |
| Revenue share | Yes. Agreed % of downstream revenue, fixed term. | None. Floor + usage only. |
| Commitment | Named design-partner agreement, joint roadmap, quarterly reviews. | Annual contract, standard SLAs, self-serve expansion. |
| Billing evidence | Signed, tamper-evident cost ledger emitted by the runtime — verifiable on your side without breaking tenant federation. | |
Why this shape: design partners take real risk with us early, so we take revenue-share risk with them - no platform bill until their agents earn. At GA we drop the revenue share and switch to a transparent floor + two usage counters.
Illustrative: at a 20% share, a partner earning $1M through MiddyMind-powered agents returns $200k — growing as the substrate compounds.
Rejected: per-seat (punishes the “substrate fades” outcome), per-connector (taxes the wedge), pure per-token (positions us as a model reseller).
Three sequenced phases: convert the signed sovereign pipeline, template it, then open the standard PaaS with a federated loop.
Ship in-country production deployments through EmpowerReg's regional partners.
Package each jurisdiction as an eval-backed template so new regulators deploy in days.
Convert design partners to floor-plus-usage pricing.
Sovereign-grade memory + governance, with a signed nation-state pipeline via EmpowerReg. Three regional partners — Network Partners (US), Excopan (Turkey), Mandala (LATAM) — are already fronting device regulators, hospital systems, and ministries of health in a $450B+ safety-intelligence market.
Written for two reviewers: VC partners underwriting a defensible substrate, and cloud program leads underwriting sovereign / in-country hyperscaler consumption. Traction, sovereign channels, market, timing, and sequenced GTM below.
Each metric below states its as-of date, an explicit definition of what counts, and a target with a confidence band. Nothing here is a projection dressed as a fact.
Confidence · High = signed / observed · Medium = active pipeline · Low = plan of record only.
EmpowerReg is a US-based global health safety intelligence company already selling into device makers and hospitals through an AI-native SaaS platform for safety, risk intelligence, regulatory / compliance workflows, and post-market surveillance. MiddyMind is embedded as the memory + governance substrate under that stack — which means EmpowerReg's three regional partners double as MiddyMind's sovereign-tenant channel from day one.
FDA-facing device makers + hospital systems.
Turkish Medicines and Medical Devices Agency (TİTCK) ecosystem.
ANVISA (Brazil), COFEPRIS (Mexico), INVIMA (Colombia) and adjacent authorities.
Ministries and regulators are standing up in-country AI mandates now.
Durable value moves to memory, governance, and eval underneath.
Federated learning on graded corrections is shippable today.
Funds the engineering to convert the signed sovereign / design-partner pipeline into.
Underwrites in-country tenant infrastructure across every EmpowerReg regional.
The compounding argument: agents are replaceable, memory isn’t. Every source a customer connects, every correction they grade, every entity they resolve makes the substrate more valuable - to them and to every partner on it. That’s the asset both VC and cloud capital are underwriting.
If you’re an investor, a design partner shipping agents into a regulated tenant, or a program office evaluating in-tenant AI infrastructure — we’d like to hear from you. Everything else can wait until we’re out of stealth.
Every intro reply includes a signed traces bundle keyed to your deployment posture. Standard reply window: 24 hours.