[GH-ISSUE #74] [Enhancement]: Memory storage question #44

New Issue

2026-06-06T22:08:14-04:00

yindo commented

2026-06-06 22:08:14 -04:00

Originally created by @adriaanvermaak on GitHub (Oct 28, 2025).
Original GitHub issue: https://github.com/vxcontrol/pentagi/issues/74

Originally assigned to: @asdek on GitHub.

Target Component

Memory System (Vector Store/Knowledge Base)

Enhancement Description

HI Team,

what i have noticed is that when a flow starts for the first on a endpoint, the memorist does produce vector stores and seems to adding all related chunks into pg.

could it be that when all flows are removed or deleted from the UI, it also cleans up the vectors? S
in the case when the first endpoint is entered for testing again with a blank slate, it basically starts the vector storage again.

perhaps some memory management to keep data indefinitely ? and add a clean up handler for maintenance ?

thank you again

Technical Details

No response

Designs and Mockups

No response

Alternative Solutions

No response

Verification

I have checked that this enhancement hasn't been already proposed
This enhancement aligns with PentAGI's goal of autonomous penetration testing
I have considered the security implications of this enhancement
I have provided clear use cases and benefits

Originally created by @adriaanvermaak on GitHub (Oct 28, 2025). Original GitHub issue: https://github.com/vxcontrol/pentagi/issues/74 Originally assigned to: @asdek on GitHub. ### Target Component Memory System (Vector Store/Knowledge Base) ### Enhancement Description HI Team, what i have noticed is that when a flow starts for the first on a endpoint, the memorist does produce vector stores and seems to adding all related chunks into pg. could it be that when all flows are removed or deleted from the UI, it also cleans up the vectors? S in the case when the first endpoint is entered for testing again with a blank slate, it basically starts the vector storage again. perhaps some memory management to keep data indefinitely ? and add a clean up handler for maintenance ? thank you again ### Technical Details _No response_ ### Designs and Mockups _No response_ ### Alternative Solutions _No response_ ### Verification - [ ] I have checked that this enhancement hasn't been already proposed - [ ] This enhancement aligns with PentAGI's goal of autonomous penetration testing - [ ] I have considered the security implications of this enhancement - [ ] I have provided clear use cases and benefits

yindo added the enhancement label 2026-06-06 22:08:14 -04:00

yindo commented

2026-06-06 22:08:15 -04:00

@asdek commented on GitHub (Oct 28, 2025):

Hello @adriaanvermaak,

Thank you for the excellent observation!

We're actually working on a memory management page that will be added in the very near future. The agent memory is divided into two types: memory tied to a specific flow and memory that exists beyond it.

For now, you can use a workaround by adding instructions to your task prompt so that the agent saves discovered results about the target as guides and checks at the beginning if a guide exists for testing that particular target. Alternatively, you can modify the system prompts for the necessary agents (e.g., for the pentester). This should work for your scenario.

Currently, the following document types are available between different runs: guides (for assistant, installer, pentester agents), code (for assistant, coder agents), and answer (for assistant and searcher agents).

Initially, flows were intended to be used for grouping by attack targets, but practice has shown that there are situations when you need to run several flows at different time intervals, so we'll try to add better management for this soon. To resume a previously stopped flow, you can first create a new assistant in it, and it will become available for continuing automation and creating new tasks. It will have the same access to files as before, but the Docker container will be created from scratch, so previously installed software won't be there.

We're also preparing a graph-based knowledge base for everything that happens during testing. For this, we'll be using our fork of the Graphiti tool with our custom taxonomy.

@asdek commented on GitHub (Oct 28, 2025): Hello @adriaanvermaak, Thank you for the excellent observation! We're actually working on a memory management page that will be added in the very near future. The agent memory is divided into two types: memory tied to a specific flow and memory that exists beyond it. For now, you can use a workaround by adding instructions to your task prompt so that the agent saves discovered results about the target as `guides` and checks at the beginning if a `guide` exists for testing that particular target. Alternatively, you can modify the system prompts for the necessary agents (e.g., for the pentester). This should work for your scenario. Currently, the following document types are available between different runs: `guides` (for assistant, installer, pentester agents), `code` (for assistant, coder agents), and `answer` (for assistant and searcher agents). Initially, flows were intended to be used for grouping by attack targets, but practice has shown that there are situations when you need to run several flows at different time intervals, so we'll try to add better management for this soon. To resume a previously stopped flow, you can first create a new assistant in it, and it will become available for continuing automation and creating new tasks. It will have the same access to files as before, but the Docker container will be created from scratch, so previously installed software won't be there. We're also preparing a graph-based knowledge base for everything that happens during testing. For this, we'll be using our [fork of the Graphiti tool](https://github.com/vxcontrol/pentagi-graphiti) with our custom [taxonomy](https://github.com/vxcontrol/pentagi-taxonomy).

yindo referenced this issue

2026-06-06 22:09:21 -04:00

[PR #44] [MERGED] Assistant mode with AI agents refactoring, streaming responses and performance boost #135

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: vxcontrol/pentagi#44