Google OKF + Hermes Agent + GBrain: Turn Any Folder Into a Knowledge Graph for M&E Professionals

Your organisation probably has the evidence already. The problem is that it is scattered across evaluation reports, indicator files, theories of change, dashboards, survey documentation, meeting notes and staff memory. This tutorial shows how to think about Google’s Open Knowledge Format (OKF), Hermes Agent and GBrain as a practical architecture for making that knowledge more structured, connected and usable by AI—without losing the human judgement that makes M&E credible.

Important: OKF is an open knowledge format, while Hermes and GBrain are separate agent/memory projects. This tutorial presents them as complementary building blocks for experimentation. They are separate projects, not one official integrated product, and their suitability for sensitive development-sector data must be assessed by your organisation.

1. What are we actually building?

Think of the stack as three layers:

The result you are aiming for is not simply “an AI that has read our folder.” It is closer to:

M&E evidence folder
        ↓
structured knowledge
        ↓
OKF-style concepts + metadata
        ↓
linked knowledge
        ↓
GBrain memory / graph / retrieval
        ↓
Hermes or another AI agent
        ↓
evidence-aware questions and workflows
        ↓
human M&E judgement

2. Why this matters for M&E

Most M&E knowledge is relational.

An indicator is related to an outcome. The outcome is related to a theory of change. A finding is related to an evaluation question. A recommendation is related to a management response. A management response is related to an action. An action is related to a follow-up review.

Yet those relationships are usually hidden inside documents.

This is the shift from document retrieval toward evidence navigation.

3. Start with a real M&E folder

Do not begin by building a huge enterprise knowledge graph. Pick one programme, one evaluation portfolio or one evidence domain.

A realistic pilot folder might contain:

• Theory of Change and results framework
• indicator reference sheets
• monitoring reports
• baseline and endline documentation
• evaluation reports
• assessment reports
• learning briefs
• management responses
• approved lessons learned
• relevant methodological guidance

Leave raw personally identifiable information, safeguarding records and other sensitive material outside the pilot until your organisation has confirmed an appropriate data-governance approach.

4. The most important concept: context

Suppose the folder contains an Excel file called:

A human who built the programme may know exactly what it means.

An AI agent does not.

It needs the surrounding context:

• What does Outcome 03 represent?
• Which indicator is stored in the file?
• What is the denominator?
• What population is included?
• Which reporting period?
• Has the indicator definition changed?
• What are the known data-quality limitations?

OKF is designed around exactly this type of contextual knowledge: metadata, context and curated insight surrounding data and systems. Google’s specification describes OKF as a minimal, human- and agent-friendly format using Markdown files with YAML frontmatter. Google Cloud’s introduction to OKF.

5. Turn an indicator into an AI-readable knowledge object

Instead of relying on the spreadsheet alone, create a small Markdown concept describing the indicator.

---
type: Indicator
title: Percentage of target households with improved income
description: Outcome indicator for household economic resilience.
tags: [income, livelihoods, outcome, resilience]
---
# Definition
Percentage of eligible target households reporting income
above the programme-defined threshold.
# Result level
Outcome
# Calculation
Eligible households above threshold /
eligible surveyed households × 100.
# Data source
Annual household survey.
# Frequency
Annual.
# Disaggregation
Sex of household head, location, disability status.
# Known limitations
Self-reported income, seasonal variation,
and potential recall bias.
# Related concepts
[Programme](../programmes/livelihood-resilience.md)
[Outcome](../outcomes/economic-resilience.md)
[Survey](../data-sources/annual-household-survey.md)

Now the indicator is not just a column in a spreadsheet. It is a piece of documented knowledge with relationships.

6. Why YAML frontmatter matters

OKF uses YAML frontmatter to provide structured metadata while leaving the body of the document flexible. The current specification defines fields including type, title, description, resource, tags and trust-related metadata. See the current OKF specification.

For M&E, useful metadata can include:

7. Build the folder around M&E concepts, not file types

A useful pilot structure could look like this:

/programme-a/
  index.md
  /programmes/
    programme-a.md
  /theories-of-change/
    programme-a-toc.md
  /outcomes/
    economic-resilience.md
  /indicators/
    indicator-001.md
    indicator-002.md
  /evaluations/
    evaluation-2026-01.md
  /findings/
    finding-001.md
  /recommendations/
    recommendation-001.md
  /decisions/
    decision-001.md
  /lessons/
    lesson-001.md
  /methods/
    sampling.md
    qualitative-analysis.md
  /data-sources/
    household-survey.md
  /context/
    country-context.md

The exact taxonomy is up to your organisation. OKF deliberately does not prescribe a fixed domain taxonomy. Its purpose is to provide a lightweight structure that can be read and exchanged across systems.

8. Create the links that make the graph useful

The graph emerges from relationships.

Programme
   ↓
Theory of Change
   ↓
Outcome
   ↓
Indicator
   ↓
Data source
   ↓
Evaluation
   ↓
Finding
   ↓
Recommendation
   ↓
Management response
   ↓
Decision
   ↓
Follow-up

For an evaluator, this relationship chain is often more valuable than a simple list of documents.

9. Where GBrain fits

GBrain is designed as a memory layer for AI agents. Its project describes a Markdown-first system of record, a self-wiring knowledge graph, hybrid search and agent-facing tools. Explore GBrain on GitHub.

For an M&E use case, the important idea is not the brand name. It is the workflow:

1. Store structured knowledge in human-readable files.
2. Create relationships between concepts.
3. Index and retrieve related knowledge.
4. Let the agent traverse connections rather than relying only on keyword search.
5. Keep the original source and human review in the loop.

10. Where Hermes Agent fits

Hermes is the agent layer that can work with tools, files and persistent workflows. The official project describes it as a self-improving AI agent with capabilities for skills, memory and searching past conversations. View Hermes Agent documentation and code.

For an M&E professional, imagine asking the agent:

The agent is useful because it can operate on the knowledge layer. The knowledge layer is useful because it gives the agent something more reliable than an unstructured folder.

11. The M&E use case: “What do we know?”

One of the most useful workflows is an evidence briefing.

Ask:

A responsible workflow should produce something like:

1. Relevant evidence — evaluations, assessments and monitoring evidence.
2. Findings — what each source actually concluded.
3. Context — where and for whom the findings apply.
4. Contradictions — where evidence disagrees.
5. Gaps — what the evidence does not answer.
6. Provenance — where each claim came from.
7. Human judgement — what the evaluator still needs to determine.

12. The M&E use case: indicator intelligence

Connect indicator definitions to the data sources and reporting rules that explain them.

Then an agent can potentially answer questions such as:

• “Which indicators changed definition during the programme?”
• “Which indicators have known data-quality limitations?”
• “Which outcome indicators are currently below target?”
• “Which indicators have no recent data?”
• “Which indicators rely on self-reported information?”

But do not let the agent silently decide what “below target” means. Your organisation needs explicit definitions and approved business rules.

13. The M&E use case: evaluation synthesis

Instead of asking AI to summarise one report, create a connected evidence base.

For example:

14. The M&E use case: recommendations that do not disappear

One of the most practical knowledge-graph applications is recommendation tracking.

Evaluation finding
      ↓
Recommendation
      ↓
Management response
      ↓
Accepted / rejected / modified
      ↓
Action
      ↓
Owner
      ↓
Deadline
      ↓
Follow-up evidence
      ↓
Status

That turns evaluation from a document-production exercise into a traceable learning cycle.

15. The M&E use case: organisational memory

Ask your organisation:

Today, the answer may live in someone’s inbox.

With structured knowledge, the relationship could be:

Data-quality problem → Evaluation recommendation → Management decision → New definition → Implementation date

That is institutional memory made navigable.

16. The M&E use case: theory-of-change checking

AI should not be allowed to declare that a theory of change is “correct.” But it can help surface relationships that deserve human scrutiny.

For example:

• Which assumptions have no supporting evidence?
• Which causal links have not been tested?
• Which outcomes have no corresponding indicators?
• Which indicators do not actually measure the stated outcome?
• Which evaluation findings challenge the current theory?

The agent becomes a question generator and evidence navigator, not the evaluator.

17. The humanitarian M&E use case

Humanitarian teams may have particularly strong reasons to structure knowledge because evidence is generated rapidly and often across multiple organisations, locations and reporting cycles.

Useful concepts might include:

• needs assessments;
• rapid monitoring;
• community feedback;
• protection concerns;
• market monitoring;
• post-distribution monitoring;
• response adaptations;
• after-action reviews.

However, humanitarian evidence often includes sensitive information. A knowledge-graph experiment should begin with appropriately sanitised or synthetic material and clear access controls.

18. Gender, equity and inclusion: do not flatten the evidence

An AI system can easily turn diverse evidence into a smooth summary.

That can be dangerous.

Make sure your knowledge objects retain distinctions such as:

• sex and gender;
• age groups;
• disability status;
• geography;
• displacement status;
• language;
• socioeconomic group;
• other contextually relevant groups.

Do not allow “the programme worked” to become the default summary when the evidence shows different effects for different populations.

19. Local knowledge needs provenance too

A knowledge graph should not automatically privilege formal donor reports over community evidence.

Instead, record the nature of each evidence source.

20. Do not let the graph manufacture consensus

This is a major M&E risk.

If an AI-generated summary is saved as a new knowledge object and later retrieved alongside the original evaluation, the summary may begin to look like an independent source.

This is one reason provenance, trust and lifecycle matter in the current OKF specification. Read Google’s discussion of OKF v0.2 trust signals.

21. A practical “knowledge status” model

For an M&E pilot, use simple lifecycle labels:

22. How to turn a normal folder into an OKF-style bundle

Start small.

Step 1 — choose one programme.

Do not index your entire organisation on day one.

Step 2 — identify 20–50 high-value concepts.

Prioritise indicators, outcomes, evaluations, findings, recommendations and key decisions.

Step 3 — create Markdown records.

Give each concept a clear title, type, source, context and relevant relationships.

Step 4 — add links.

Connect concepts using Markdown links and consistent naming.

Step 5 — create an index.

Make it easy for people and agents to discover the bundle.

Step 6 — validate.

Check broken links, duplicate concepts, missing provenance and outdated definitions.

Step 7 — introduce retrieval.

Only after the knowledge layer is useful should you add an agent workflow.

23. A simple pilot architecture

Existing M&E folder
        ↓
Curate & de-duplicate
        ↓
Create OKF-style knowledge bundle
        ↓
Validate provenance + lifecycle
        ↓
GBrain / retrieval layer
        ↓
Hermes Agent
        ↓
Evidence questions
        ↓
Human verification
        ↓
Approved learning / decision

This architecture keeps an important boundary: AI can help retrieve and synthesise; people remain responsible for evaluation judgement and organisational decisions.

24. What should the agent never silently do?

• Invent an indicator definition.
• Merge different versions of a result framework.
• Treat an AI summary as an independent evidence source.
• Erase uncertainty from evaluation findings.
• Turn correlation into causation.
• Expose confidential participant information.
• Treat a historical programme decision as current policy.
• Convert a community comment into a representative finding without qualification.

25. How M&E teams can use Hermes + GBrain safely

For an initial experiment, keep the scope deliberately narrow.

Before putting sensitive information into any third-party or cloud-based workflow, check your organisation’s data-protection, contractual, safeguarding and information-security requirements.

26. A 30-minute M&E knowledge-graph exercise

1. 5 minutes: choose one evaluation or programme folder.
2. 5 minutes: identify five important concepts.
3. 10 minutes: draw the relationships between them.
4. 5 minutes: identify what information is missing.
5. 5 minutes: ask what an AI agent could do if those relationships were explicit.

For example:

That small exercise often reveals where your organisation’s institutional memory is actually stored.

27. What success looks like

Do not judge the pilot by how impressive the AI conversation looks.

Measure whether it improves evidence work.

• Retrieval: Can staff find the right evidence faster?
• Traceability: Can important claims be traced to sources?
• Currency: Can users distinguish current from historical knowledge?
• Coverage: Are important evidence gaps visible?
• Learning: Can previous recommendations be connected to subsequent decisions?
• Quality: Does AI-assisted synthesis preserve uncertainty and context?
• Use: Does the system actually help programme teams make better decisions?

28. The bigger lesson for M&E

The interesting part of OKF is not simply that it uses Markdown.

The important idea is that knowledge can be structured so that both humans and AI systems can navigate it. Google’s current specification describes OKF as intentionally minimal, portable and human- and agent-readable, with version control and trust-related metadata built into the newer specification. Read the specification.

For M&E, this suggests a useful direction:

29. Key questions for an M&E team discussion

1. What knowledge would we want a new M&E officer to understand on their first day?
2. Where is that knowledge currently stored?
3. Which indicators have ambiguous or outdated definitions?
4. Which evaluation recommendations have disappeared into reports?
5. Which decisions depend on institutional memory?
6. Which evidence should never be exposed to an AI agent?
7. What would we gain if our evidence relationships were explicit?

30. A practical M&E workflow: from folder to evidence graph

Here is a useful sequence for an M&E team that wants to test the approach without turning it into a major IT project.

1. Curate: choose a bounded folder and remove obvious duplicates, drafts and material that should not enter the pilot.
2. Describe: turn high-value concepts—programmes, outcomes, indicators, evaluations, findings and recommendations—into Markdown records with clear metadata.
3. Connect: link related concepts so the evidence trail becomes navigable.
4. Verify: record provenance, review status and lifecycle. Do not treat AI-generated text as validated simply because it is fluent.
5. Retrieve: use GBrain or another suitable retrieval layer to search and traverse the knowledge.
6. Act: use Hermes or another agent to perform bounded tasks such as evidence briefs, recommendation tracking or gap analysis.
7. Review: require an M&E professional to check important claims, interpretations and decisions against the underlying sources.

30.1. M&E readiness checklist

31. Final takeaway

OKF gives you a way to structure knowledge. GBrain can provide a memory and graph layer. Hermes can provide an agent interface.

But the technology is only one part of the system.

The difficult work is deciding:

• what counts as evidence;
• what counts as validated knowledge;
• which sources take precedence;
• how uncertainty is represented;
• who can access sensitive information;
• who is accountable for interpretation.

That is why this is an M&E challenge as much as an AI challenge.

Further reading and tools

Open Knowledge Format specification

Google Cloud: Introducing OKF

Google Cloud: OKF v0.2 and trust signals

Hermes Agent project

GBrain project

EvalCommunity Academy · Practical AI, evidence, monitoring, evaluation, learning and research resources for M&E, MEL/MEAL, humanitarian and international development professionals.