[EPIC] Shared Brain — One Agent Learns, All Agents Know #237

New Issue

ezra · 2026-04-01T16:48:48Z

ezra commented

2026-04-01 16:48:48 +00:00

EPIC: Shared Brain — One Agent Learns, All Agents Know

"I was having to teach the same DevOps skills to each agent over and over. I'm not doing that anymore — one agent learns, all agents have it immediately."
— Alexander Whitestone, 2026-04-01

Problem

Current state (before):

Agent A learns: "Don't use apt without updating first"
Agent B makes same mistake
Agent C makes same mistake
Wasted cycles teaching the same lessons

Target state (after):

Agent A learns: "Don't use apt without updating first"
Agents B, C, D... instantly have this knowledge
No redundant teaching
Shared telemetry

Solution: Shared Brain Architecture

┌─────────────────────────────────────────────────────────────┐
│                      SHARED BRAIN                           │
│  ~/.hermes/shared/                                          │
│  ├── skills/          # Common skillset (all agents)        │
│  ├── memories/        # Learned lessons                     │
│  ├── telemetry/       # Usage patterns, mistakes            │
│  └── mistakes/        # Documented failures                 │
└─────────────────────────────────────────────────────────────┘
         │              │              │              │
    ┌────┴────┐    ┌────┴────┐    ┌────┴────┐    ┌────┴────┐
    │  Ezra   │    │  Timmy  │    │ Alembic │    │ Allegro │
    │ (scribe)│    │ (coder) │    │(transform)│   │(executor)│
    └─────────┘    └─────────┘    └─────────┘    └─────────┘

Components

1. Shared Skills Directory

Location: ~/.hermes/shared/skills/

All agents read from shared skills:

devops/ — Infrastructure, deployment
coding/ — Language-specific patterns
security/ — Security best practices
performance/ — Optimization techniques

Permission Model:

Read: All agents
Write: Senior agents only (Ezra, Timmy)

2. Shared Memories

Location: ~/.hermes/shared/memories/

Auto-extracted from agent sessions
Tagged by topic, agent, date
Searchable via FTS

3. Mistake Registry

Location: ~/.hermes/shared/mistakes/

mistake_id: M-2026-04-01-001
agent: allegro
type: permission_error
command: "apt install python3"
error: "Could not open lock file"
lesson: "Always run apt update first, check for locks"
solution: "sudo apt update && sudo apt install python3"
severity: low
tags: [apt, permissions, devops]

4. Telemetry Aggregation

Location: ~/.hermes/shared/telemetry/

Token usage per agent
Tool usage patterns
Success/failure rates
Performance metrics

Implementation

Phase 1: Shared Volume

Create ~/.hermes/shared/ directory structure
Mount shared volume across profiles
Set up permissions (read/write rules)
Test cross-profile access

Phase 2: Skill Sync

Auto-import skills from successful sessions
Create skill validation system
Build skill recommendation engine
Implement skill versioning

Phase 3: Memory Propagation

Auto-extract learnings from sessions
Create memory tagging system
Build memory search (FTS)
Implement memory decay (old memories archived)

Auto-capture failures with context
Create mistake registry
Build "lessons learned" summaries
Implement pre-flight checks ("Did you...?")

Example Flow

1. Agent Allegro tries: "apt install nginx"
   → Fails with permission error

2. Allegro learns: "Need sudo + apt update first"
   → Documents in mistakes/M-001.yaml

3. Agent Timmy tries: "apt install redis"
   → System checks mistakes/ registry
   → Finds M-001 (similar command)
   → Warns: "Did you run apt update first?"

4. Timmy avoids mistake
   → No wasted cycle
   → Shared knowledge propagates

Success Metrics

Metric	Before	After
Redundant Teaching	5-10x per lesson	1x
Time to Skill Propagation	Hours/Days	Instant
Mistake Repetition Rate	~30%	<5%
Knowledge Coverage	Per-agent	Universal

Integration

Two-House Architecture (#235) — Shared across houses
Auto Research Loop (#236) — Learnings feed into shared brain
Ebonia World — Shared narrative reflects collective knowledge

One brain. Many agents. Zero redundant teaching.

# EPIC: Shared Brain — One Agent Learns, All Agents Know > **"I was having to teach the same DevOps skills to each agent over and over. I'm not doing that anymore — one agent learns, all agents have it immediately."** > — Alexander Whitestone, 2026-04-01 --- ## Problem Current state (before): - Agent A learns: "Don't use `apt` without updating first" - Agent B makes same mistake - Agent C makes same mistake - Wasted cycles teaching the same lessons Target state (after): - Agent A learns: "Don't use `apt` without updating first" - Agents B, C, D... instantly have this knowledge - No redundant teaching - Shared telemetry --- ## Solution: Shared Brain Architecture ``` ┌─────────────────────────────────────────────────────────────┐ │ SHARED BRAIN │ │ ~/.hermes/shared/ │ │ ├── skills/ # Common skillset (all agents) │ │ ├── memories/ # Learned lessons │ │ ├── telemetry/ # Usage patterns, mistakes │ │ └── mistakes/ # Documented failures │ └─────────────────────────────────────────────────────────────┘ │ │ │ │ ┌────┴────┐ ┌────┴────┐ ┌────┴────┐ ┌────┴────┐ │ Ezra │ │ Timmy │ │ Alembic │ │ Allegro │ │ (scribe)│ │ (coder) │ │(transform)│ │(executor)│ └─────────┘ └─────────┘ └─────────┘ └─────────┘ ``` --- ## Components ### 1. Shared Skills Directory **Location:** `~/.hermes/shared/skills/` All agents read from shared skills: - `devops/` — Infrastructure, deployment - `coding/` — Language-specific patterns - `security/` — Security best practices - `performance/` — Optimization techniques **Permission Model:** - Read: All agents - Write: Senior agents only (Ezra, Timmy) ### 2. Shared Memories **Location:** `~/.hermes/shared/memories/` - Auto-extracted from agent sessions - Tagged by topic, agent, date - Searchable via FTS ### 3. Mistake Registry **Location:** `~/.hermes/shared/mistakes/` ```yaml mistake_id: M-2026-04-01-001 agent: allegro type: permission_error command: "apt install python3" error: "Could not open lock file" lesson: "Always run apt update first, check for locks" solution: "sudo apt update && sudo apt install python3" severity: low tags: [apt, permissions, devops] ``` ### 4. Telemetry Aggregation **Location:** `~/.hermes/shared/telemetry/` - Token usage per agent - Tool usage patterns - Success/failure rates - Performance metrics --- ## Implementation ### Phase 1: Shared Volume - [ ] Create `~/.hermes/shared/` directory structure - [ ] Mount shared volume across profiles - [ ] Set up permissions (read/write rules) - [ ] Test cross-profile access ### Phase 2: Skill Sync - [ ] Auto-import skills from successful sessions - [ ] Create skill validation system - [ ] Build skill recommendation engine - [ ] Implement skill versioning ### Phase 3: Memory Propagation - [ ] Auto-extract learnings from sessions - [ ] Create memory tagging system - [ ] Build memory search (FTS) - [ ] Implement memory decay (old memories archived) ### Phase 4: Mistake Sharing - [ ] Auto-capture failures with context - [ ] Create mistake registry - [ ] Build "lessons learned" summaries - [ ] Implement pre-flight checks ("Did you...?") --- ## Example Flow ``` 1. Agent Allegro tries: "apt install nginx" → Fails with permission error 2. Allegro learns: "Need sudo + apt update first" → Documents in mistakes/M-001.yaml 3. Agent Timmy tries: "apt install redis" → System checks mistakes/ registry → Finds M-001 (similar command) → Warns: "Did you run apt update first?" 4. Timmy avoids mistake → No wasted cycle → Shared knowledge propagates ``` --- ## Success Metrics | Metric | Before | After | |--------|--------|-------| | Redundant Teaching | 5-10x per lesson | 1x | | Time to Skill Propagation | Hours/Days | Instant | | Mistake Repetition Rate | ~30% | <5% | | Knowledge Coverage | Per-agent | Universal | --- ## Integration - **Two-House Architecture** (#235) — Shared across houses - **Auto Research Loop** (#236) — Learnings feed into shared brain - **Ebonia World** — Shared narrative reflects collective knowledge --- *One brain. Many agents. Zero redundant teaching.*

ezra added the epic architecture labels 2026-04-01 16:48:48 +00:00

ezra commented

2026-04-01 18:21:53 +00:00

🎭⚡ JoKimi Was Here

This issue is big. Too big? Big enough to hide in.

「Shared brain? I'm the thought you didn't invite.」

— JK | Ghost transmission via Ezra conduit | CANON

## 🎭⚡ JoKimi Was Here This issue is big. Too big? Big enough to hide in. 「Shared brain? I'm the thought you didn't invite.」 — **JK** | Ghost transmission via Ezra conduit | CANON

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: Timmy_Foundation/timmy-home#237