Edit: github-backup/docs/12-metaswarm-gbrain-inspiration.md

Note Content (Markdown)

# Metaswarm and GBrain Inspiration

## Summary

Adopt the operating principles, not the whole stack immediately.

Metaswarm is most useful as a model for disciplined agent work:

- written specifications;
- role ownership;
- review gates;
- adversarial review;
- independent validation;
- persistent task state;
- selective knowledge priming;
- closure learning.

GBrain is most useful as a model for a brain daemon:

- Markdown as source material;
- hybrid search plus synthesis;
- graph links;
- gap analysis;
- recurring dream-cycle maintenance;
- skill packs;
- company-brain access control;
- MCP exposure to multiple clients.

Neither should be dropped onto the live n8n VM before Easier has a memory
filter, benchmark set and connector policy.

## Metaswarm Patterns to Adopt

Source: `https://github.com/dsifry/metaswarm`

### Written Work Contracts

Every material agent task should have:

- goal;
- definition of done;
- allowed files/systems;
- expected output;
- review path.

For Easier, this maps to:

- agent role briefs;
- SOP improvement requests;
- connector activation reviews;
- weekly operating reviews;
- future Easier Now implementation tasks.

### Review Gates

Metaswarm uses specialist review gates and independent validation. Easier
should use a lighter agency version:

- COO review for role/spec quality;
- security/access review before new connectors;
- relationship review before client-visible workflows;
- finance/legal review before spend, pricing, billing or contracts.

### Selective Priming

Metaswarm's knowledge base grows, but agents load only relevant entries. Easier
should copy this strongly:

- agent receives `SOUL.md`;
- agent receives its role file;
- agent receives current priority page;
- agent retrieves only relevant SOPs/evidence;
- agent does not load the whole vault by default.

### Closure Learning

Every meaningful run should end with:

- what worked;
- what failed;
- what the user corrected;
- what should become a rule, SOP, benchmark or gate;
- what should be forgotten.

This is the "self-learning organisation" loop from the 8 Figure guide, but
kept grounded in reviewable files.

## GBrain Patterns to Adopt

Sources:

- `https://github.com/garrytan/gbrain`
- `https://raw.githubusercontent.com/garrytan/gbrain/master/INSTALL_FOR_AGENTS.md`

### Search Plus Synthesis

Raw search is not enough. The user needs answers with citations and a clear
statement of what the brain does not know.

Adopt this evaluation target even before adopting gbrain itself:

- retrieve evidence;
- synthesise answer;
- cite sources;
- state gaps and staleness.

### Graph and Entity Awareness

Easier has many relationship-heavy workflows: clients, prospects, partners,
team members, offers, tools and products.

The vault should make those entities first-class:

- people;
- organisations;
- clients/prospects;
- projects;
- offers;
- tools;
- agents;
- decisions.

### Dream Cycle, But With a Filter

GBrain's recurring maintenance pattern is attractive, but Easier should define
the filter before running it:

- dedupe people/company pages;
- surface contradictions;
- fix citations;
- identify stale notes;
- propose knowledge promotions;
- never bulk-import raw sources into working memory.

### Cost and Mode Awareness

GBrain's agent install protocol explicitly asks the operator to choose search
mode because retrieval can create surprise cost. Easier should apply that same
principle to all agent loops:

- daily reviews are capped;
- weekly reviews are capped;
- high-context retrieval requires explicit run mode;
- expensive synthesis jobs show expected cost where possible.

## What to Delay

Delay:

- installing gbrain on the live n8n VM;
- running local embedding/model workloads on the small VM;
- broad Slack import;
- automated email/calendar ingestion;
- autonomous dream cycles over private data;
- multi-agent job queues before one COO loop works.

Reason:

- the n8n VM already hosts production-ish services;
- disk and RAM are limited;
- live connectors raise privacy and reputational risk;
- context quality matters more than agent count.

## Benchmark Before Adoption

Test Hermes `llm-wiki`, QMD and gbrain against the same Easier benchmark:

- exact source recall;
- fuzzy founder recall;
- client relationship prep;
- SOP lookup;
- contradiction handling;
- temporal change detection;
- safe refusal.

Adopt the simplest layer that passes.