The House Stack
For each system an agent reaches, the product AgentMesh runs behind it, why, and what your organization can choose instead.
two stacks
House Stack and Your Stack
An agent reaches systems outside its own folder: a database, a place for files, a model to think with, a voice. A package names the system and never the product. The stack says which product stands behind each system.
The House Stack is AgentMesh's own choices, one list for everybody. Your Stack is your organization's choices, made on the Stack page in the app. Where your organization has chosen nothing for a system, the House Stack fills in. Where neither has anything, an agent that needs the system is not installed until you choose one, and the app says so in words before anything is built.
A deploy states what the agent will use rather than asking: Database, Supabase (Your Stack) · Agent memory, Hindsight (House Stack), with a Change link. It asks only twice: once per organization, the first time an agent would keep your own data (database, files or agent memory) in the House Stack, and whenever neither stack provides a system the agent needs.
the list
What the House Stack runs
Read off what AgentMesh actually runs on 2026-09-27. Where the House Stack cannot cover a system today, the row says so rather than pretending.
| System | House product | Why | You can choose instead |
|---|---|---|---|
| Text reasoning | Google Gemini | Every harness AgentMesh deploys runs on Gemini unless its owner names another model, and the key sits on the machine that deploys it. | Anthropic, Google (your own key), OpenRouter |
| Image reading | Google Gemini | The same Gemini models read images, so one provider covers text and pictures. | Anthropic, Google (your own key), OpenRouter |
| Speech to text | Google Cloud Speech-to-Text | The transcript step in AgentMesh's own agents already calls it, under the same Google account as the models. | Google (your own key) |
| Text to speech | Google Gemini speech | AgentMesh's narration uses Gemini's voices, so recordings sound the same across agents. | Google (your own key), ElevenLabs |
| Draft host | Google Cloud Storage | Drafts are written to AgentMesh's own bucket and served under agentmesh.ai/draft, so a reviewer opens them with a plain link. | Google Cloud Storage, Amazon S3 (recorded; the shared file script cannot sign S3 requests yet) |
| Activity record | NATS JetStream | The record of what agents did is kept in the same JetStream that carries their messages, so nothing is recorded twice. | Nothing. It is AgentMesh's own record of the network. |
| Agent memory | Hindsight | An open source memory service (MIT licensed) that keeps what an agent learns and can recall and reflect on it. AgentMesh runs one shared copy with its own database, and each agent reaches only its own memory. Covered once the service is switched on in a deployment; until then the Stack page says it is not switched on. | Nothing yet. Something else can be recorded, and a deploy then needs it bound by hand. |
| Database | Not covered yet | AgentMesh does not run a database for agents. | Supabase, Firestore |
| Files | Not covered yet | AgentMesh does not run a file store for agents. | Google Cloud Storage, Amazon S3 (recorded, not yet usable) |
| Sign-in | Not covered yet | AgentMesh signs its own people in with email links and passkeys, and does not offer that sign-in to agents. | Okta, Google |
| Not covered yet | AgentMesh sends its own mail through Resend, and does not lend that sender to agents. | Resend, Amazon SES | |
| MCP gateway | Not covered yet | AgentMesh runs an MCP connector for its own tools, and does not offer it as a gateway in front of your MCP servers. | Your own gateway |
| API gateway | Only yours | An API gateway fronts your own services, so only your organization can run it. | Your own gateway |
"Not covered yet" is a statement about today. When AgentMesh starts running one of these for agents, the row changes and agents in organizations that chose nothing pick it up at their next deploy.
agent memory
How agent memory works
An agent gets three tools: remember,
recall and reflect.
Each call is signed with the agent's own key and goes to AgentMesh's memory
gateway, which maps the agent to its own bank from its registered name and
forwards the call to Hindsight. An agent cannot name another agent's bank,
list banks, delete memories or change a bank's settings; the gateway refuses
all of those.
remember works in the background: Hindsight
extracts facts with a model, which takes a few seconds on long input, so a
recall straight after may not find it yet. A bank is made when an agent is
installed and deleted with the agent. The prompts Hindsight sends to its model
carry memory text, and AgentMesh runs it set to keep no copies of them.
switching products
Moving data between products
Choosing a different product for a system never moves data. What your agents stored under the old product stays there, and at their next deploy they start empty in the new one. The Stack page says so before you save. To bring the data along, copy it across before the agents' next deploy:
| From, to | How |
|---|---|
| Supabase to another Supabase project | pg_dump the schema your agents write to from the old project's session pooler, and psql it into the new one. |
| Supabase and Firestore | Export the tables or collections as JSON and write them back through the new product's own import. Row ids are kept, because the shared store script upserts on them. |
| Google Cloud Storage to another bucket | gcloud storage rsync -r gs://old/prefix gs://new/prefix, keeping the prefix the agents write under. |
| Agent memory, House Stack to your own | No export is offered through the memory gateway yet. Ask AgentMesh for a copy of an agent's bank before you switch. |