Legal & Advice

Project Memory Agent

Whoever happens to be in the room knows whether we have built something like this before. Otherwise the search begins.

Draws on architecture decisions and lessons learned to answer what has been built before.

Open agent
Public administration

Four processes that take up the most time today

Developed in Germany · Hosted in the EU

Developed in Germany · Hosted in the EU or on-premises

Delivered together with

STACKIT Sovereign hosting in Germany, single-tenant from the first user.
Rödl Legal, tax and compliance expertise for legally sound deployment.
4P Consulting Advisory and implementation for the public sector, starting with the AI readiness assessment.

One question during the bid phase, two days of searching through old project folders.

The request comes from sales, in the middle of the bid phase. Who gets asked depends on who is available. The senior consultant names two projects off the top of their head; the rest sits in decision papers, final reports and lessons learned across three separate repositories.

3repositoriesDecision papers are in the project repository, final reports in document management, lessons learned in Confluence.
62projects since 2019That is how many projects could be considered for a comparison at all.
4colleagues askedAnyone who finds nothing asks around, and every query costs the person asked a search of their own.
2 daysto an answerBy the time the 2022 decision paper is on the table, the bid phase is half over.

With hits, decision and reference in a single answer, search effort falls by up to 60%.

The reduction of up to 60% in search effort is this agent's headline figure. The number of repositories, the project portfolio and the two days to an answer describe a typical sequence at a consultancy; none of it was measured.

Three steps through the process.

The senior consultant weighs up whether an old decision fits the new project; the agent finds the comparable projects and shows the decision taken at the time alongside them, with its reference.

Run the request against the project portfolio

The agent breaks the question down into attributes such as client type, architecture and volume. It uses them to search decision papers, final reports and lessons learned.

Request on bid A-4471, 62 projects searched, 5 comparable ones found

Show decisions with references

For every hit, the agent names the architecture decision taken at the time and where it is recorded. It lists rejected variants separately.

Decision paper for Project Nordsee 2022, section 4.2, page 11

Hand the hits over for bid work

The agent compiles hits, decisions and lessons learned in one overview with sources. The senior consultant chooses what goes into the bid.

projektgedaechtnis_A-4471.pdf, 5 hits, 12 references

What the agent does not do.

It does not carry an old decision over to the new project.

It shows what was decided on Project Nordsee 2022 and why. The project manager clarifies whether that holds here.

It does not judge the quality of a final report.

If the lessons learned contain only one sentence, it reproduces that one sentence with its reference. The knowledge manager decides what follows from it.

It does not rank the five hits by similarity.

For each project it gives the year, the type of client and the reference for the decision, for example Project Nordsee 2022, section 4.2, page 11. The senior consultant decides which case comes closest to bid A-4471.

Connected systems

  • Confluence
  • document management system
  • project repository

Operations

  • Deployment modelEU cloud, private cloud or on-premises
  • Data processingGDPR-compliant, no model training on your data
  • Responsible teamsProject management, architecture, sales

The questions that come up before the first meeting.

How quickly is the Project Memory Agent ready to use?

A pilot typically runs on your own data within 4 to 6 weeks. The agent connects to your existing systems rather than running as a separate, parallel environment.

Where is the data processed?

EU cloud, private cloud or on-premises. Processing is GDPR-compliant and models are never trained on your data. Details in the Trust Center.

Where does the figure “up to 60% less time spent searching” come from?

It is based on experience from comparable projects and process benchmarks, not a commitment. We work out the realistic effect for your organisation during the assessment before the pilot.

Related agents.

Intake review, case 2026/04711

Skip to content

Page footer