keendreams-security

<p> <picture> <source media="(prefers-color-scheme: dark) and (max-width: 600px)" srcset="docs/brand/hero-mobile-dark.svg"> <source media="(max-width: 600px)" srcset="docs/brand/hero-mobile-light.svg"> <source media="(prefers-color-scheme: dark)" srcset="docs/brand/hero-dark.svg"> <img src="docs/brand/hero-light.svg" width="1280" alt="KeenDreams Security Memory by Agent9. Security decisions, remembered with evidence."> </picture> </p> <p align="left"> <a href="https://github.com/Agent9AI/keendreams-security/actions/workflows/ci.yml"><img src="https://github.com/Agent9AI/keendreams-security/actions/workflows/ci.yml/badge.svg" alt="CI"></a> <a href="https://github.com/Agent9AI/keendreams-security/actions/workflows/codeql.yml"><img src="https://github.com/Agent9AI/keendreams-security/actions/workflows/codeql.yml/badge.svg" alt="CodeQL"></a> <a href="https://github.com/Agent9AI/keendreams-security/actions/workflows/secret-scan.yml"><img src="https://github.com/Agent9AI/keendreams-security/actions/workflows/secret-scan.yml/badge.svg" alt="Secret scan"></a> <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-235d42" alt="MIT license"></a> </p>

Give your security applications and agents a memory they can explain.

KeenDreams Security is an open-source security memory layer for applications and agents. Record evidence, review proposed facts, and retrieve your team's decisions through semantic vector and keyword search. Default recall includes only trusted, currently valid facts, with their evidence and history. Deploy it in your own Cloudflare account and connect through MCP.

No Tenable account or Tenable One license is required to use KeenDreams Security. The optional Tenable Hexa workflow connects exposure data with that memory. See the live results and what remains to verify.

Start here  ·  Product tour  ·  Deploy  ·  Architecture  ·  Tool reference  ·  Tenable Hexa  ·  Documentation

| Your first step | What you get | |---|---| | Try it without accounts | Inspect sample evidence, review one proposal, and verify the audit chain locally. | | Connect your application or agent | Deploy your own memory server and connect through OAuth-protected MCP. | | Evaluate Tenable Hexa | Read scoped exposure data and test it against your team's recorded decisions. |

Listed on CyberAgents Exchange · What is verified · Get help

<a id="the-product-in-two-pages"></a>

Review the evidence

See what supports the claim before making the decision. The review queue places the proposed relationship, its source, the quoted evidence, and the reviewer's actions together.

<p> <a href="docs/product-tour.md#review-queue"> <picture> <source media="(max-width: 600px)" srcset="docs/images/review-detail-mobile.png"> <img src="docs/images/review-detail.png" width="1280" alt="Actual review component with a proposed vulnerability, its quoted evidence, and Confirm and Reject controls"> </picture> </a> </p>

_{Review queue detail. Captured from the local demo with sample evidence.}

Inspect the history behind the answer. Administrators can verify the audit chain and roll decisions back to an earlier entry. The rollback is itself recorded.

<p> <a href="docs/product-tour.md#administration"> <picture> <source media="(max-width: 600px)" srcset="docs/images/audit-detail-mobile.png"> <img src="docs/images/audit-detail.png" width="1280" alt="Actual administration component showing the verified audit chain and rollback controls"> </picture> </a> </p>

_{Administration detail. Captured from the same local demo. <a href="docs/product-tour.md">Open the full product tour</a>.}

<details> <summary>Watch the review queue turn proposals into decisions</summary>

The sequence ends on the admin page, where the recorded decisions are checked against the hash chain. Walk through the actual tool responses.

</details>

For a hands-on evaluation, run the local demo with sample data on your own computer, or follow the deployment guide to host KeenDreams Security in your Cloudflare account.

<a id="evidence-first-trust-is-a-separate-decision"></a>

Trust is a separate decision

<p> <picture> <source media="(prefers-color-scheme: dark) and (max-width: 600px)" srcset="docs/brand/trust-path-mobile-dark.svg"> <source media="(max-width: 600px)" srcset="docs/brand/trust-path-mobile-light.svg"> <source media="(prefers-color-scheme: dark)" srcset="docs/brand/trust-path-dark.svg"> <img src="docs/brand/trust-path-light.svg" width="1280" alt="An unconfirmed claim is excluded from default recall. Browser review can make it trusted and eligible for recall while valid. Explicit admin allowlists are a separate trust path."> </picture> </p>

1. Record raw evidence as a quoted episode. Credential-shaped strings are redacted.
2. Propose a relationship that cites that episode. New facts extracted by a model start unconfirmed.
3. Review in a signed-in browser. There is no MCP tool that confirms a proposal.
4. Recall trusted, currently valid facts with their evidence and decision history.

Administrators can separately allowlist a specific automation identity and source to write trusted facts. A caller cannot grant itself that permission or choose a trusted status. Read the trust model.

For the ordinary MCP path, the result is observable: a newly proposed fact is absent from default recall; after browser confirmation, it appears with its provenance. See the reproducible walkthrough.

What your team gets back

| The recurring problem | The useful answer | |---|---| | A finding was already investigated | Retrieve the earlier decision and its supporting evidence. | | A risk acceptance outlived its intended window | An expiry and reason are required; expired acceptances stop appearing as current. | | A previously fixed vulnerability returns | A newly trusted finding supersedes the remediation, preserving the timeline. | | An auditor asks who decided and why | Inspect the actor, time, evidence, and history of the relationship. | | One service is unavailable | Recall falls back to keyword search when the vector or embedding service fails. |

Explore the workflows.

Extend an application or an agent

A security application can use KeenDreams to retrieve earlier findings and decisions during an investigation. An MCP-connected agent can record evidence, propose relationships, and recall reviewed knowledge in a later session. Both use the same memory tools and trust rules; neither gains authority to approve a finding simply by connecting.

Origin Observatory is the example security application for the planned integration walkthrough. It has its own investigation tools and evidence records. The intended demonstration shows Origin displaying a Tenable finding, recording supporting evidence in KeenDreams, and retrieving a reviewed decision in a later investigation. Origin is a separate Agent9 project, not a dependency or part of this repository's runtime. Its connector belongs in a separate repo.

That application walkthrough has not yet been completed or photographed as live integration proof. The product screenshots above show the existing KeenDreams local demo. An agent-focused walkthrough will be documented separately. Application and agent integration model.

<a id="how-it-runs"></a>

Your account. Your memory.

<p> <picture> <source media="(prefers-color-scheme: dark) and (max-width: 600px)" srcset="docs/brand/architecture-mobile-dark.svg"> <source media="(max-width: 600px)" srcset="docs/brand/architecture-mobile-light.svg"> <source media="(prefers-color-scheme: dark)" srcset="docs/brand/architecture-dark.svg"> <img src="docs/brand/architecture-light.svg" width="1280" alt="MCP clients and browser reviewers connect through the Worker and OAuth. Cloudflare Access handles sign-in. Each client has a separate Durable Object and SQLite memory, its own Vectorize namespace, and access to Workers AI."> </picture> </p>

Each client gets a Durable Object with SQLite and full-text search. Vectorize adds semantic recall in a separate namespace for that client. Workers AI supplies embeddings and optional suggestions. A registry controls access and trusted sources.

| Boundary | Implementation | |---|---| | Identity | Cloudflare Access for SaaS, OAuth 2.1, PKCE, and dynamic client registration | | Memory | Per-client SQLite, evidence-linked relationships, validity windows, and history | | Retrieval | Keyword + vector search, followed by trust and validity filtering | | Human control | Browser review, administrative allowlists, and audited rollback | | Ownership | Your Worker, account, data, model configuration, and billing |

Architecture and diagrams · Security model and data flows

Tools

Nine MCP tools. Six reads, three writes. All four behavior hints are explicit.

| Read | Purpose | Write | Purpose | |---|---|---|---| | recall | Ask in plain language | record_episode | Store quoted evidence | | find_facts | Filter exact relationships | assert_fact | Propose an evidence-linked fact | | get_entity | Inspect one entity | suggest_facts | Propose relationships from an episode | | explore_graph | Follow trusted relationships | | | | fact_history | Inspect the full timeline | | | | list_proposals | Inspect the review queue | | |

Full tool reference, output examples, and annotations.

Deploy in your account

The deployment guide covers prerequisites, bindings, settings, and connecting an MCP client. The deployer owns the infrastructure; there is no service operated by Agent9 that receives your evidence. Cloudflare hosting and selected AI services have their own requirements and costs.

Sign-in setup

Configure Cloudflare Access for SaaS and the Worker secrets, then connect your MCP client. Follow the sign-in instructions.

Using it with Tenable

Tenable adds current exposure evidence; KeenDreams makes reviewed knowledge reusable across security applications and agents. Through Hexa MCP, a connected client can retrieve scoped asset and vulnerability information, compare it with prior decisions, and record new evidence-backed proposals. Where returned by the selected Tenable tools, prioritization, affected-asset details, and remediation information can give the reviewer more context than a finding title alone.

The included /hexa-to-memory skill coordinates that workflow using approved read tools. See Hexa to Memory in action: a live session that recognized a confirmed accepted risk, with its reviewer and expiry. Tenable access remains read-only; memory writes happen in KeenDreams. A decision recorded here does not create a Tenable risk-acceptance rule or change a scan result. Hexa requires an eligible Tenable One license, enabled MCP acc