monobrain

Deploy 60+ specialized agents in coordinated swarms with self-learning capabilities, fault-tolerant consensus, and enterprise-grade security.

Why Monobrain? Claude Code is powerful — but it thinks alone. Monobrain gives it a brain trust: a coordinated swarm of 100+ specialized agents that share memory, reach consensus, learn from every task, and route work to the right specialist automatically. Built on a WASM-powered intelligence layer, it gets smarter every session.

How Monobrain Works

User → Monobrain (CLI/MCP) → Router → Swarm → Agents → Memory → LLM Providers
                           ↑                          ↓
                           └──── Learning Loop ←──────┘

📐 Expanded Architecture — Full system diagram with RuVector intelligence

flowchart TB
    subgraph USER["👤 User Layer"]
        U[User]
        CC[Claude Code]
    end

    subgraph ENTRY["🚪 Entry Layer"]
        CLI[CLI / MCP Server]
        AID[AIDefence Security]
    end

    subgraph ROUTING["🧭 Routing Layer"]
        KW[Keyword Pre-Filter]
        SEM[Semantic Router]
        LLM_FB[LLM Fallback · Haiku]
        TRIG[MicroAgent Triggers]
        HK[17 Hooks · Event Bus]
    end

    subgraph SWARM["🐝 Swarm Coordination"]
        TOPO[Topologies<br/>hierarchical/mesh/adaptive]
        CONS[Consensus<br/>Raft/BFT/Gossip/CRDT]
        CLM[Claims<br/>Trust Tiers]
        GOV[Guidance<br/>Policy Gates]
    end

    subgraph AGENTS["🤖 100+ Agents"]
        AG1[coder]
        AG2[tester · reviewer]
        AG3[architect · planner]
        AG4[security-auditor]
        AG5[devops · sre]
        AG6[60+ more...]
    end

    subgraph RESOURCES["📦 Resources"]
        MEM[(AgentDB · HNSW · SQLite)]
        PROV[Providers<br/>Claude/GPT/Gemini/Ollama]
        WORK[12 Workers<br/>ultralearn/audit/optimize]
    end

    subgraph RUVECTOR["🧠 RuVector Intelligence"]
        direction TB
        SONA[SONA<br/>Self-Optimize<br/>&lt;0.05ms]
        EWC[EWC++<br/>Anti-Forgetting]
        FLASH[Flash Attention<br/>2.49–7.47x]
        HNSW_I[HNSW<br/>150x–12500x]
        RB[ReasoningBank<br/>RETRIEVE→JUDGE→DISTILL]
        LORA[LoRA/MicroLoRA<br/>128x compress]
    end

    subgraph LEARNING["🔄 Learning Loop"]
        L1[RETRIEVE] --> L2[JUDGE] --> L3[DISTILL] --> L4[CONSOLIDATE] --> L5[ROUTE]
    end

    U --> CC --> CLI --> AID
    AID --> KW & SEM & LLM_FB & TRIG & HK
    KW & SEM & LLM_FB & TRIG --> TOPO & CONS & CLM & GOV
    TOPO & CONS --> AG1 & AG2 & AG3 & AG4 & AG5 & AG6
    AG1 & AG2 & AG3 & AG4 & AG5 & AG6 --> MEM & PROV & WORK
    MEM --> SONA & EWC & FLASH & HNSW_I & RB & LORA
    LORA --> L1
    L5 -.->|loop| SEM

RuVector Intelligence Components:

Component	Purpose	Performance
SONA	Self-Optimizing Neural Architecture — learns optimal routing	<0.05ms adaptation
EWC++	Elastic Weight Consolidation — prevents catastrophic forgetting	Preserves all learned patterns
Flash Attention	Optimized attention computation	2.49–7.47× speedup
HNSW	Hierarchical Navigable Small World vector search	150×–12,500× faster
ReasoningBank	Pattern storage with RETRIEVE→JUDGE→DISTILL pipeline	Sub-ms recall
Hyperbolic	Poincaré ball embeddings for hierarchical data	Better code relationship mapping
LoRA / MicroLoRA	Low-Rank Adaptation weight compression	128× compression ratio
Int8 Quantization	Memory-efficient weight storage	~4× memory reduction
9 RL Algorithms	Q-Learning, SARSA, A2C, PPO, DQN, A3C, TD3, SAC, HER	Task-specific policy learning

Get Started Fast

Option 1 — npx (recommended):

npx monobrain@latest init --wizard
claude mcp add monobrain -- npx -y monobrain@latest mcp start
npx monobrain@latest daemon start
npx monobrain@latest doctor --fix

Option 2 — Clone from GitHub:

git clone https://github.com/nokhodian/monobrain.git
cd monobrain
npm install
node packages/@monobrain/cli/bin/cli.js init --wizard

# Wire up the MCP server in Claude Code
claude mcp add monobrain -- node "$PWD/packages/@monobrain/cli/bin/cli.js" mcp start

New to Monobrain? You don't need to learn 170+ MCP tools or 41 CLI commands up front. After running init, just use Claude Code normally — the hooks system automatically routes tasks to the right agents, learns from successful patterns, and coordinates multi-agent work in the background.

Key Capabilities

🤖 100+ Specialized Agents — Ready-to-use AI agents for every engineering domain: coding, review, testing, security, DevOps, mobile, ML, blockchain, SRE, and more. Each optimized for its specific role.

🐝 Coordinated Agent Swarms — Agents organize into teams using hierarchical (queen/workers) or mesh (peer-to-peer) topologies. They share context, divide work, and reach consensus — even when agents fail.

🧠 Learns From Every Session — Successful patterns are stored in HNSW-indexed vector memory and reused. Similar tasks route to the best-performing agents automatically. Gets smarter over time without retraining.

⚡ 3-Tier Cost Routing — Simple transforms run in WASM at <1ms and $0. Medium tasks use Haiku. Complex reasoning uses Sonnet/Opus. Smart routing cuts API costs by 30–50%.

🔌 Deep Claude Code Integration — 170+ MCP tools expose the full platform directly inside Claude Code sessions. The hooks system fires on every file edit, command, task start/end, and session event.

🔒 Production-Grade Security — CVE-hardened AIDefence layer blocks prompt injection, path traversal, command injection, and credential leakage. Per-agent WASM/Docker sandboxing with cryptographic audit proofs.

🧩 Extensible Plugin System — Add custom capabilities with the plugin SDK. Distribute via the IPFS-based decentralized marketplace. 20 plugins available today across core, integration, optimization, and domain categories.

🏛️ Runtime Governance — @monobrain/guidance compiles your CLAUDE.md into enforced policy gates: destructive-op blocking, tool allowlists, diff size limits, secret detection, trust tiers, and HMAC-chained proof envelopes.

Claude Code: With vs Without Monobrain

Capability	Claude Code Alone	Claude Code + Monobrain
Agent Collaboration	One agent, isolated context	Swarms with shared memory and consensus
Hive Mind	⛔ Not available	Queen-led hierarchical swarms with 3+ queen types
Consensus	⛔ No multi-agent decisions	Byzantine fault-tolerant (f < n/3), Raft, Gossip, CRDT
Memory	Session-only, ephemeral	HNSW vector memory + knowledge graph, persistent cross-session
Self-Learning	Static, starts fresh every time	SONA self-optimization, EWC++ anti-forgetting, pattern reuse
Task Routing	Manual agent selection	Intelligent 3-layer routing (keyword → semantic → LLM), 89% accuracy
Simple Transforms	Full LLM call every time	Agent Booster (WASM): <1ms, $0 cost
Background Work	Nothing runs automatically	12 workers auto-dispatch on hooks events
LLM Providers	Anthropic only	Claude, GPT, Gemini, Cohere, Ollama with failover and cost routing
Security	Standard Claude sandboxing	CVE-hardened, WASM/Docker sandbox per agent, cryptographic proofs
Governance	CLAUDE.md is advisory	Runtime-enforced policy gates with HMAC audit trail
Cost	Full LLM cost every task	30–50% reduction via WASM, caching, smart routing

Architecture Deep Dives

🧭 Intelligent Task Routing — 3-layer pipeline that routes every request

Every request passes through a 3-layer pipeline before any agent sees it:

Request
  │
  ├─► [Layer 1] Keyword pre-filter     → instant match, zero LLM cost
  │
  ├─► [Layer 2] Semantic routing       → embedding similarity vs. agent catalog
  │
  └─► [Layer 3] LLM fallback (Haiku)  → Haiku-powered classification for ambiguous tasks

Once classified, the task hits the 3-tier cost model:

Tier	Handler	Latency	Cost	Used for
1	Agent Booster (WASM)	<1ms	$0	Simple transforms (var→const, add types, logging)
2	Haiku	~500ms	~$0.0002	Moderate tasks, summaries, Q&A
3	Sonnet / Opus	2–5s	$0.003–$0.015	Architecture, security, complex reasoning

Hook signals — what the system emits to guide routing:

# Agent Booster can handle it — skip LLM entirely
[AGENT_BOOSTER_AVAILABLE] Intent: var-to-const
→ Use Edit tool directly, <1ms, $0

# Model recommendation for Task tool
[TASK_MODEL_RECOMMENDATION] Use model="haiku" (complexity=22)
→ Pass model="haiku" to Task tool for cost savings

Microagent trigger scanner — 10 specialist agents with keyword frontmatter triggers:

Domain	Trigger keywords	Agent
Security	auth, injection, CVE, secret	security-architect
DevOps	deploy, CI/CD, pipeline, k8s	devops-automator
Database	query, schema, migration, index	database-optimizer
Frontend	React, CSS, component, SSR	frontend-dev
Solidity	contract, ERC, Solidity, DeFi	solidity-engineer

🐝 Swarm Coordination — How agents organize and reach consensus

Agents organize into swarms with configurable topologies and consensus algorithms:

Topology	Best for	Consensus
Hierarchical	Coding tasks, feature work (default)	Raft (leader-based)
Mesh	Distributed exploration, research	Gossip / CRDT
Adaptive	Auto-switches based on load	Byzantine (BFT)

Consensus algorithms:

Algorithm	Fault tolerance	Use case
Raft	f < n/2	Authoritative state, coding swarms
Byzantine (BFT)	f < n/3	Untrusted environments
Gossip	Eventual consistency	Large swarms (100+ agents)
CRDT	No coordination overhead	Conflict-free concurrent writes

Anti-drift swarm configuration (recommended for all coding tasks):

npx monobrain@latest swarm init \
  --topology hierarchical \
  --max-agents 8 \
  --strategy specialized \
  --consensus raft

Setting	Why it prevents drift
hierarchical	Coordinator validates every output against the goal
max-agents 6–8	Smaller team = less coordination overhead
specialized	Clear roles, no task overlap
raft	Single leader maintains authoritative state

Task → agent routing:

Task	Agents
Bug fix	coordinator · researcher · coder · tester
New feature	coordinator · architect · coder · tester · reviewer
Refactor	coordinator · architect · coder · reviewer
Performance	coordinator · perf-engineer · coder
Security audit	coordinator · security-architect · auditor

🧠 Self-Learning Intelligence — How Monobrain gets smarter every session

Every task feeds the 4-step RETRIEVE-JUDGE-DISTILL-CONSOLIDATE pipeline:

RETRIEVE  ──►  JUDGE  ──►  DISTILL  ──►  CONSOLIDATE
   │               │            │               │
HNSW search   success/fail   LoRA extract   EWC++ preserve
150x faster    verdicts       128x compress  anti-forgetting

Memory architecture:

Feature	Details
Episodic memory	Full task histories with timestamps and outcomes
Entity extraction	Automatic extraction of code entities into structured records
Procedural memory	Learned skills from .monobrain/skills.jsonl
Vector search	384-dim embeddings, sub-ms retrieval via HNSW
Knowledge graph	PageRank + community detection for structural insights
Agent isolation	Per-agent memory scopes prevent cross-contamination
Hybrid backend	SQLite + AgentDB, zero native binary dependencies

Specialization scorer — per-agent, per-task-type success/failure tracking with time-decay. Feeds routing quality over time. Persists to .monobrain/scores.jsonl.

⚡ Agent Booster (WASM) — Skip the LLM for simple code transforms

Agent Booster uses WebAssembly to handle deterministic code transforms without any LLM call:

Intent	Example	vs LLM
var-to-const	var x = 1 → const x = 1	352× faster
add-types	Add TypeScript annotations	420× faster
add-error-handling	Wrap in try/catch	380× faster
async-await	.then() → async/await	290× faster
add-logging	Insert structured debug logs	352× faster
remove-console	Strip all console.* calls	352× faster
format-string	Modernize to template literals	400× faster
null-check	Add ?. / ?? operators	310× faster

When hooks emit [AGENT_BOOSTER_AVAILABLE], Claude Code intercepts and uses the Edit tool directly — zero LLM round-trip.

💰 Token Optimizer — 30–50% API cost reduction

Smart caching and routing stack multiplicatively to reduce API costs:

Optimization	Savings	Mechanism
ReasoningBank retrieval	–32%	Fetches relevant patterns, not full context
Agent Booster transforms	–15%	Simple edits skip LLM entirely
Pattern cache (95% hit rate)	–10%	Reuses embeddings and routing decisions
Optimal batch size	–20%	Groups related operations
Combined	30–50%	Multiplicative stacking

🏛️ Governance — Runtime policy enforcement from CLAUDE.md

@monobrain/guidance compiles CLAUDE.md into a 7-phase runtime enforcement pipeline:

CLAUDE.md ──► Compile ──► Retrieve ──► Enforce ──► Trust ──► Prove ──► Defend ──► Evolve

Phase	Enforces
Enforce	Destructive ops, tool allowlist, diff size limits, secret detection
Trust	Per-agent trust accumulation with privilege tiers
Prove	HMAC-SHA256 hash-chained audit envelopes
Defend	Prompt injection, memory poisoning, collusion detection
Evolve	Policy drift detection, auto-update proposals

1,331 tests · 27 subpath exports · WASM security kernel

Quick Start

Prerequisites

• Node.js 20+ (required)
• Claude Code — npm install -g @anthropic-ai/claude-code

Installation

One-line (recommended):

curl -fsSL https://cdn.jsdelivr.net/gh/nokhodian/monobrain@main/scripts/install.sh | bash

Via npx:

npx monobrain@latest init --wizard

Manual:

# Register MCP server with Claude Code
claude mcp add monobrain -- npx -y monobrain@latest mcp start

# Start background worker daemon
npx monobrain@latest daemon start

# Health check
npx monobrain@latest doctor --fix

First Commands

# Spawn an agent
npx monobrain@latest agent spawn -t coder --name my-coder

# Launch a full swarm
npx monobrain@latest hive-mind spawn "Refactor auth module to use OAuth2"

# Search learned patterns
npx monobrain@latest memory search -q "authentication patterns"

# Dual Claude + Codex workflow
npx monobrain-codex dual run feature --task "Add rate limiting middleware"

⚡ Slash Commands

Type these directly in Claude Code. No setup beyond npx monobrain init.

🤖 Agent Intelligence

Command	What it does
/specialagent	Scores all 60+ agents against your task and picks the best one (or recommends a full swarm config). Prevents wasting a generic coder on a job that needs a Database Optimizer or tdd-london-swarm.
/use-agent [slug]	Instantly activates a non-dev specialist agent — Sales Coach, TikTok Strategist, Legal Compliance Checker, UX Researcher, etc. Without a slug, auto-picks from conversation context.
/list-agents [category]	Lists all available specialist agents, optionally filtered by category (marketing, sales, design, academic, product, project-management, support).

🌐 Browser & UI Automation

Command	What it does
/ui-test <url>	Full UI test run: opens the URL, snapshots interactive elements, walks golden-path flows, tests edge cases, reports pass/fail/warn. Powered by agent-browser.
/browse <url>	Navigates to a URL and walks through it step by step — narrating what's on screen, proposing actions, and helping you accomplish tasks via the browser.
/crawl <url>	Crawls a website — extracts links, text, structured data, or anything you specify. Great for scraping, auditing, or data extraction tasks.
/browser	Raw agent-browser session: opens an interactive browser automation context with snapshot, click, fill, and screenshot tools. Use when you need fine-grained control.

🏛️ SPARC Methodology

Command	What it does
/sparc	Runs the full SPARC orchestrator — breaks down your objective, delegates to the right modes, and coordinates the full development lifecycle.
/sparc spec-pseudocode	Captures requirements, edge cases, and constraints, then translates them into structured pseudocode ready for implementation.
/sparc code	Auto-coder mode — writes clean, efficient, modular code from pseudocode or a spec.
/sparc debug	Debugger mode — traces runtime bugs, logic errors, and integration failures systematically.
/sparc security-review	Security reviewer — static and dynamic audit, flags secrets, poor module boundaries, and injection risks.
/sparc devops	DevOps mode — CI/CD, Docker, deployment automation.
/sparc docs-writer	Writes clear, modular Markdown documentation: READMEs, API references, usage guides.
/sparc refinement-optimization-mode	Refactors, modularizes, and improves system performance. Enforces file size limits and dependency hygiene.
/sparc integration	System integrator — merges outputs of all modes into a working, tested, production-ready system.

🐝 Swarm & Memory

Command	What it does
/monobrain-swarm	Coordinates a multi-agent swarm for complex tasks — spawns agents, distributes work, waits for results, synthesizes output.
/monobrain-memory	Interacts with the AgentDB memory system — store, search, retrieve, and inspect patterns across sessions.
/monobrain-help	Shows all Monobrain CLI commands and usage reference inline.

Pro tip — automatic activation: You don't need to type slash commands for most flows. The UserPromptSubmit hook reads every prompt and automatically suggests the right slash command (or activates it) based on what you wrote. /specialagent activates when you ask "which agent", /ui-test activates when you say "test the UI", /browse activates when you say "go to the website", etc.

Agents

100+ specialized agents across every engineering domain:

🔧 Core Development

Agent	Specialization
coder	Clean, efficient implementation across any language
reviewer	Code review — correctness, security, maintainability
tester	TDD, integration, E2E, coverage analysis
planner	Task decomposition, sprint planning, roadmap
researcher	Deep research, information gathering
architect	System design, DDD, architectural patterns
analyst	Code quality analysis and improvement

🔒 Security

Agent	Specialization
security-architect	Threat modeling, secure design, vulnerability assessment
security-auditor	Smart contract audits, CVE analysis
security-engineer	Application security, OWASP, secure code review
threat-detection	SIEM rules, MITRE ATT&CK, detection engineering
compliance-auditor	SOC 2, ISO 27001, HIPAA, PCI-DSS

🐝 Swarm & Consensus

Agent	Specialization
hierarchical-coordinator	Queen-led coordination with specialized worker delegation
mesh-coordinator	P2P mesh, distributed decision-making, fault tolerance
adaptive-coordinator	Dynamic topology switching, self-organizing
byzantine-coordinator	BFT consensus, malicious actor detection
raft-manager	Raft protocol, leader election, log replication
gossip-coordinator	Gossip-based eventual consistency
crdt-synchronizer	Conflict-free replication
consensus-coordinator	Sublinear solvers, fast agreement

🚀 DevOps & Infrastructure

Agent	Specialization
devops-automator	CI/CD pipelines, infrastructure automation
cicd-engineer	GitHub Actions, pipeline creation
sre	SLOs, error budgets, chaos engineering
incident-response	Production incident management, post-mortems
database-optimizer	Schema design, query optimization, PostgreSQL/MySQL
data-engineer	Data pipelines, lakehouse, dbt, Spark, streaming

🌐 Frontend, Mobile & Specialized

Agent	Specialization
frontend-dev	React/Vue/Angular, UI, performance optimization
mobile-dev	React Native iOS/Android, cross-platform
accessibility	WCAG, screen readers, inclusive design
solidity-engineer	EVM smart contracts, gas optimization, DeFi, L2
ml-engineer	ML model development, training, deployment
embedded-firmware	ESP32, STM32, FreeRTOS, Zephyr, bare-metal
backend-architect	Scalable systems, microservices, API design
technical-writer	Developer docs, API references, tutorials

🔀 GitHub Workflow Automation

Agent	Specialization
pr-manager	PR lifecycle, review coordination, merge management
code-review-swarm	Parallel multi-agent code review
release-manager	Automated release coordination, changelog
repo-architect	Repository structure, multi-repo management
issue-tracker	Issue management, project coordination
workflow-automation	GitHub Actions creation and optimization

🔬 SPARC Methodology

Agent	Specialization
sparc-coord	SPARC orchestrator across all 5 phases
specification	Requirements analysis and decomposition
pseudocode	Algorithm design, logic planning
architecture	System design from spec
refinement	Iterative improvement
sparc-coder	TDD-driven implementation from specs

View all: npx monobrain@latest agent list

Live Statusline

Monobrain adds a real-time six-row dashboard to Claude Code:

▊ Monobrain v1.0.0  ○ IDLE  nokhodian  │  ⎇ main  +1  ~9921 mod  ↑5  │  🤖 Sonnet 4.6
──────────────────────────────────────────────────────────────────────────────────────
💡  INTEL    ▱▱▱▱▱▱ 3%   │   📚 190 chunks   │   76 patterns
──────────────────────────────────────────────────────────────────────────────────────
🐝  SWARM    0/15 agents   ⚡ 14/14 hooks   │   🎯 3 triggers · 24 agents   │   → ROUTED  👤 Coder  81%
──────────────────────────────────────────────────────────────────────────────────────
🧩  ARCH     82/82 ADRs   │   DDD ▰▰▱▱▱ 40%   │   🛡️ ✖ NONE   │   CVE not scanned
──────────────────────────────────────────────────────────────────────────────────────
🗄️  MEMORY   0 vectors   │   2.0 MB   │   🧪 66 test files   │   MCP 1/1  DB ✔
──────────────────────────────────────────────────────────────────────────────────────
📋  CONTEXT  📄 SI 80% budget (1201/1500 chars)   │   🏗 ▰▰▱▱▱ 2/5 domains   │   💾 47 MB RAM

Row	Shows
Header	Version, session state, git user, branch, uncommitted changes
INTEL	Intelligence score, knowledge chunks indexed, learned patterns
SWARM	Active agents, hook count, microagent triggers, last routing result
ARCH	ADR compliance, DDD domain coverage, security gates, CVE status
MEMORY	Vector count, DB size, test file count, MCP/DB health
CONTEXT	Shared instructions budget, domain coverage, RAM usage

Toggle compact ↔ full: /ts — Full reference: tagline.md

Dual-Mode Collaboration

Run Claude Code and OpenAI Codex workers in parallel with shared memory:

# Pre-built templates
npx monobrain-codex dual run feature --task "Add OAuth authentication"
npx monobrain-codex dual run security --target "./src"
npx monobrain-codex dual run bugfix --task "Fix race condition in session handler"

# Custom pipeline
npx monobrain-codex dual run \
  --worker "claude:architect:Design the API contract" \
  --worker "codex:coder:Implement the endpoints" \
  --worker "claude:tester:Write integration tests" \
  --worker "codex:optimizer:Reduce allocations"

Worker dependency order: Architect (L0) → Coder + Tester (L1) → Reviewer (L2) → Optimizer (L3)

Template	Workers	Pipeline
feature	Architect → Coder → Tester → Reviewer	Full feature development
security	Analyst → Scanner → Reporter	Security audit
refactor	Architect → Refactorer → Tester	Code modernization
bugfix	Researcher → Coder → Tester	Bug investigation and fix

Packages

todo: write about packages of this app

Plugins

todo: write about plugins of this app 20 plugins via the IPFS-distributed registry:

npx monobrain@latest plugins list
npx monobrain@latest plugins install @monobrain/plugin-name
npx monobrain@latest plugins create my-plugin

Contributing

git clone https://github.com/nokhodian/monobrain.git
cd monobrain/packages
pnpm install
pnpm test

Support

Documentation	github.com/nokhodian/monobrain
Issues	github.com/nokhodian/monobrain/issues
Enterprise	monoes.me

MIT — nokhodian

Acknowledgements

Monobrain builds on ideas, patterns, and research from the following projects:

Repository	What we took
ruvnet/ruflo	Original skeleton — swarm coordination, hooks system, and SPARC methodology
msitarzewski/agency-agents	Agent architecture patterns and multi-agent md files
microsoft/autogen	Human oversight interrupt gates, AutoBuild ephemeral agents, procedural skill learning from executions, and tool-retry patterns
crewAIInc/crewAI	Multi-tier memory (short/long/entity/contextual), role/goal/backstory agent registry, task context chaining, and output schema patterns
langchain-ai/langgraph	Graph checkpointing + resume, StateGraph workflow DSL (fan-out/fan-in, conditional, loops), and entity extraction from conversation state
All-Hands-AI/OpenHands	Per-agent Docker/WASM sandboxing, semantic versioned agent registry (AgentHub), and EventStream session replay
agno-agi/agno	AgentMemory knowledge base architecture and team-level agent coordination class
huggingface/smolagents	Explicit planning step before execution and ManagedAgent delegation wrapper
pydantic/pydantic-ai	Typed Agent[Deps, Result] I/O schemas, auto-retry on validation failure, TestModel for deterministic CI, and dynamic system prompt functions
BAAI/AgentSwarm (Agency Swarm)	Declared directed communication flows between agents and shared instruction propagation
BerriAI/atomic-agents	BaseIOSchema typed agent contracts and SystemPromptContextProvider composition
stanfordnlp/dspy	BootstrapFewShot + MIPRO automatic prompt optimization pipeline
aurelio-labs/semantic-router	Utterance-based RouteLayer replacing static routing codes, dynamic routes, and hybrid routing mode
langfuse/langfuse	Unified trace/span/generation observability hierarchy, per-agent cost attribution, latency views, and prompt version management
karpathy/autoresearch	Experiment loop protocol (BASELINE/KEEP/DISCARD results.tsv), fixed time-budget per run, and Best-Fit Decreasing bin packing for API chunking — wired into @monoes/graph pipeline
safishamsi/graphify	Knowledge graph construction approach, AST-based node/edge extraction, community detection with Louvain, and GRAPH_REPORT.md report format — foundation for @monoes/graph
google/gvisor (paper)	gVisor runsc OCI-compatible runtime — reduces Docker container syscall surface from 350+ to ~50 interceptions; wired into SandboxConfig.use_gvisor and buildDockerArgs()
Indirect Injection research (follow-up)	Prompt injection via external tool content — validateExternalContent() in @monobrain/security applies pattern + optional aidefence semantic scan to all externally-sourced content
FOREVER Forgetting Curve	Exponential importance-weighted forgetting curve (importanceScore × e^(−λt)) replacing linear decay — implemented in LearningBridge.decayConfidences() and MemoryEntry.importanceScore
Awesome RLVR	Reinforcement Learning with Verifiable Rewards — hooksModelOutcome now accepts verifier_type (tsc/vitest/eslint/llm_judge) and exit_code to derive grounded binary reward signals
ERL — Experiential Reflective Learning	Structured {condition, action, confidence} heuristics extracted at hooks_post-task and injected as ranked hints into hooks_pre-task suggestions via the heuristics memory namespace
A-MEM — Agentic Memory	Zettelkasten-style automatic note linking — after every bridgeStoreEntry, top-3 HNSW neighbors above 0.7 similarity receive a similar causal edge via bridgeRecordCausalEdge
DSPy	Bayesian exploration option (bayesian: true) added to PromptOptimizer.optimize() — shuffles trace scores with U(0,0.1) noise before selectExamples to escape local optima
Collaborative Memory Promotion	Auto-promote memory access_level from private → team when 3+ distinct agents read an entry within 24 h — implemented via agent_reads table in SQLiteBackend and checkAndPromoteEntry()
Zep / Graphiti — Bi-Temporal Knowledge Graph (repo)	Separates event time T from ingestion time T' — MemoryEntry.eventAt nullable field + event_at SQLite column for temporal filtering without index rebuilds; 94.8% on Deep Memory Retrieval at 90% lower latency than MemGPT
HippoRAG 2 — PPR Graph Retrieval	Personalized PageRank over the memory reference graph — MemoryGraph.pprRerank() expands HNSW candidates one hop via MemoryEntry.references, boosting associative recall by up to 20% on multi-hop QA
RAPTOR — Recursive Abstractive Tree Indexing	Cluster episodic entries → summarize each cluster → store as contextual-tier entry — implemented in the consolidate background worker (runConsolidateWorker), creating RAPTOR's tree within existing stores
Multi-Agent Reflexion (MAR)	Heterogeneous Diagnoser → Critic×2 → Aggregator reflection loop — hooks_post-task now returns marReflection when a task fails, specifying the four agent roles and spawn order
TextGrad — Automatic Differentiation via Text (Nature)	LLM textual gradients flow backward through the pipeline — on hooks_post-task failure a textual_gradient critique is stored to the gradients memory namespace for next-prompt injection; +20% on LeetCode-Hard
CP-WBFT — Confidence-Probe Weighted BFT	Confidence-weighted voting replaces one-node-one-vote — weightedTally() in consensus/vote-signer.ts scales each agent's vote by its confidence score, tolerating 85.7% fault rate (AAAI 2026)
GraphRAG + Practical GraphRAG (Practical)	Community-level global query answering — MemoryGraph.getCommunitySummaries() returns top-k community descriptors (nodeCount, avgPageRank) for prepending to semantic search results; enables thematic reasoning over the entire knowledge base
MemPalace	Spatially-organized verbatim memory with Wing→Room→Hall hierarchy, Okapi BM25 + closet-topic hybrid retrieval, score-based L1 promotion, and temporal knowledge graph — implemented in .claude/helpers/memory-palace.cjs; injects L0 identity + L1 essential story on every session start via SessionStart hook; achieves 96.6% LongMemEval recall without summarization