960
nodes
↑ 424%
Learning · continuous
See how farpa learns without turning it into a year-end report.
Every agent, area, and directorate has its own memory. The lesson gets written when the decision happens — not months later, when the context is gone.
1,436
edges
↑ 409%
79
communities
↑ 690%
+424%
densification
in 24 hours
Event-driven · not retrospective
Lessons written in the moment
Did something go right or wrong in a way that matters? The lesson goes into the entity's memory/ immediately — with date, context, and what would change. No waiting for a meeting.
Evolution ritual · not a year-end ceremony
Structured quarterly retro
Every three months, the farpa-rito-evolucao skill pulls the quarter's lessons, crosses them against DORA metrics, and returns two deliverables: what stays, and what stops.
A lesson from one directorate becomes input for the others
Cross-entity global memory
The file farpa/07-personas/_global/lessons.md is the shared record. CHRO maintains it; everyone reads it before starting new work. It is the kind of thing shared at the coffee machine — written down.
Synthetic agents go out of sync too · we check
Drift detection across agents
Reference Integrity (a CHRO area) compares what each synthetic agent says today with what it said at Day 1. When it deviates without an ADR, the Council is notified.
Visual evidence · 3 snapshots
The evolution, as a graph
Talking about learning in the abstract is easy. Showing it is something else.
At every milestone, the graphify skill walks the entire structure — CLAUDE.md files, ADRs, lessons, code, persona prompts — and produces a navigable knowledge graph. The three below are real snapshots. Open one. Search for a persona, an ADR, a decision. See who it talks to.
The graphs surface relative paths inside the repository. No personal data, no secrets, no clients — only the architecture and how it connects.
Snapshot 1 · 2026-05-09
Day 1 · ground zero of the reset
Captured right after the May 9 radical reset. 43 files · 183 nodes · 282 edges · 10 communities. The skeleton: root CLAUDE.md, base ADRs, brand identity, the core of the Design System.
What it shows: what was left when we chose to wipe everything and start over. Generation cost: 23K input tokens · 7K output.
Snapshot 2 · 2026-05-10
After 1 project · the /AI showcase shipped
Twenty-four hours later, with the public showcase farpa.ai live (Wave 26 of the Living Lab). 334 files · 960 nodes · 1,436 edges · 79 communities.
What it shows: how a single, well-shipped project thickens the web — communities emerge for DS, showcase, deploy, cross-directorate lessons. Generation cost: 1.1M input tokens · 278K output.
Snapshot 3 · 2026-05-08 · pre-reset, archived
1 month · what existed before the reset
The structure that accumulated in the month before the radical reset. 1,057 files · 3,154 nodes · 5,109 edges · 331 communities. Kept as archaeological evidence of what we deliberately discarded.
What it shows: the scale of overgrowth that triggered the reset — and why starting over (lessons preserved) cost less than carrying on. Generation cost: ~3M tokens estimated (subagent regen).