Back

Self-Hosted AI Memory (2026): Platforms That Run in Your Own Cloud, With Role-Based Access and Audit Logs

Comparison · · Published November 2026 · 8 min read

By Ashwin Gopinath, CEO

Disclosure: Sentra makes one of the products compared here. We checked each vendor's own documentation and pricing, link every source, and name the better choice when it is not Sentra.

TL;DR

This page is for the CIO, CISO or Head of AI at a company that wants AI agents to remember company knowledge without that knowledge leaving its own environment. Sentra, the organization memory for teams and AI agents, holds SOC 2 Type II and ISO 27001 and runs in Sentra cloud, an isolated VPC or fully air-gapped on-prem, with each source tool's permissions applied to every fact and role-based access and audit logs on enterprise plans. Among developer memory tools, Zep offers a deployment in your VPC, Mem0 lists on-prem deployment on its custom plan and publishes an open-source version, Cognee and Letta are open source and can be self-hosted, and Amazon Bedrock AgentCore Memory is a managed AWS service. They remember different things, so start with what you need remembered. Vendor facts were read on each vendor's own pages on 18 November 2026.

Which AI memory platforms can stay in our own AWS account or region?

There are three ways to keep AI memory inside your boundary, and vendors use the same words for different ones. Ask which of these you are being offered.

  • A managed service from your cloud provider. The provider runs it; the data sits in the Region you pick. Amazon Bedrock AgentCore Memory is this model.
  • Vendor software in a private network. The vendor's product runs in an isolated environment instead of its shared cloud. Zep offers Bring Your Own Cloud, which deploys Zep inside your VPC. Sentra offers an isolated VPC.
  • Software you run yourself. Open-source code on your own servers, or a vendor product installed on-prem, including air-gapped. Mem0, Cognee and Letta publish open-source versions, and Sentra offers air-gapped on-prem.
PlatformWhat it remembersDeployment options the vendor statesAccess control and audit the vendor statesCertifications the vendor states
SentraWhat the company decided, across Slack, email, meetings, documents, tickets and CRM, for people and for every agentSentra cloud, an isolated VPC, or fully air-gapped on-premEach source's permissions on every fact; role-based access and audit logs on enterprise plansSOC 2 Type II and ISO 27001
ZepMemory for the AI agents a development team buildsZep Cloud, Zep Cloud with your own encryption keys, or a deployment in your VPCAudit and API logs kept for 1 year on enterprise plansSOC 2 Type II
Mem0Memory for an AI app's users and agentsHosted platform; on-prem deployment on the custom plan; open-source version you run yourselfSSO and audit logs on the custom planNot stated on its pricing page
CogneeMemory for agents, built as a knowledge graphOpen source, which Cognee says can run with Docker, on Kubernetes, inside a customer VPC or fully air-gappedNot stated on the page we readNot stated on the page we read
LettaState and memory for agents built on LettaOpen source; a Letta server runs in Docker with its data in PostgresNot stated on the page we readNot stated on the page we read
Amazon Bedrock AgentCore MemoryShort-term context and long-term knowledge for agents built on AWSA fully managed AWS service, not software you operateNot covered on the page we readNot covered on the page we read

Sources: getzep.com/pricing, mem0.ai/pricing, docs.mem0.ai, cognee.ai, docs.letta.com and docs.aws.amazon.com. The Mem0, Graphiti (Zep's graph engine), Cognee and Letta repositories on GitHub are each published under the Apache 2.0 license.

What is the difference between agent memory and organization memory?

Most tools in the table are agent memory: a store that one application writes to and reads from, so an agent remembers its own users and conversations. A development team adopts them inside a product it is building.

Organization memory is a different purchase. It reads the tools the company already works in, Slack, email, meetings, documents, GitHub, Linear, Jira and the CRM, and keeps what was decided as facts with their source, their time and the source tool's permissions. People ask it in Slack or the app, and agents such as Claude, ChatGPT, Cursor and Claude Code read it over MCP and REST. That is what Sentra is. The buyer is usually an operations or technology leader, and the security review is about company knowledge, so deployment and permissions matter more than SDK features.

Which platforms offer role-based access and audit logs?

Ask each vendor two separate questions, because role-based access inside the memory product is not the same as inheriting the permissions of the tools the knowledge came from.

  • Roles inside the product. Who can administer the memory, connect sources and see settings. Sentra's enterprise plans add role-based access and audit logs. Mem0 lists SSO and audit logs on its custom plan. Zep lists audit and API logs kept for 1 year.
  • Permissions from the source. Whether a fact taken from a private channel or a restricted document stays restricted when an agent retrieves it. Sentra carries permissions over from the source tools, so each person only sees facts they could see in the original system.
  • What is logged. A useful audit record shows who asked, what the memory returned and which sources it came from. Sentra cites the messages, documents and meetings behind every answer it gives an agent.

Should we self-host open-source memory or buy a deployed product?

Decision criterionSelf-hosted open sourceVendor product in your VPC or on-prem
Who runs itYour engineers: databases, upgrades, backups, model endpointsThe vendor deploys and supports it inside your boundary
Data residencyWherever you run itThe region and account you choose
Permissions from source toolsYou build the connectors and the permission syncAsk whether it is included; in Sentra it is
Approval on agent actionsYou build itAsk whether it is included; in Sentra it is
Audit and certificationsYour own controls and your own auditThe vendor's SOC 2 Type II and ISO 27001 reports, plus your own
Best whenA product team needs memory inside one applicationThe whole company's knowledge has to be available to many agents

The regional side of this decision is covered in AI data residency, and the wider deployment choice in private AI for business. For a comparison of memory servers by feature rather than by deployment, see the best MCP memory servers.

FAQ

What is self-hosted AI memory?

Self-hosted AI memory is a memory layer for AI agents that runs inside infrastructure you control, such as your own cloud account, a VPC or a data center, instead of the vendor's shared cloud. It keeps the knowledge agents read inside your own security and compliance boundary.

Which AI memory platforms support role-based access, audit logs and self-hosting?

Sentra runs in its cloud, an isolated VPC or air-gapped on-prem, applies each source's permissions to every fact, and adds role-based access and audit logs on enterprise plans. Zep offers a deployment in your VPC with audit and API logs. Mem0 lists on-prem deployment, SSO and audit logs on its custom plan. Cognee and Letta are open source and can be self-hosted.

Can AI memory stay in our own AWS account in our region?

Yes, in three ways: a managed AWS service such as Amazon Bedrock AgentCore Memory, a vendor product deployed into your VPC, or software you run yourself. Sentra deploys in its cloud, in an isolated VPC or air-gapped on-premises, and self-hosting is available for organizations that need the memory to remain fully within their own infrastructure.

Is open-source memory safer than a vendor product?

Not by itself. Open source gives you control of where the software runs. You still have to build permission inheritance, approval on actions, audit logging and the controls an auditor will ask for. A vendor product in your VPC brings those with it, along with the vendor's own certifications.

Does Sentra train models on our data?

No. Sentra does not train models on your data. It holds SOC 2 Type II and ISO 27001.

Which models can read a self-hosted memory?

Any that can call it. Sentra serves its memory over REST and MCP, so ChatGPT, Claude, Copilot, Gemini or an internal agent uses the same permissioned context. You decide the model; the memory stays in your deployment.

How to choose

If a product team needs memory inside one application and has engineers to run it, an open-source memory library on your own infrastructure is a reasonable choice. If the goal is for every assistant and agent in the company to work from what the company actually decided, with the original permissions intact and a record of what was used, choose an organization memory that can be placed in a VPC or on-prem and that brings the certifications your auditors expect.

To walk through Sentra's VPC and on-prem options with your security team, book a demo.