pulse-os 10.3.0

PULSE: A Neural Operating System for AI Swarms.

PULSE: The Autonomous Neural Operating System for AI Swarms

"Intelligence without memory is a waste of money." — Thursday Noon, 2026

The Problem Everyone Ignores

You run 5 AI agents. They don't share memory. They overwrite each other's work. They hallucinate into your production files. You manually route every prompt. You babysit a $200/month swarm that forgets you every session.

No framework solves this. CrewAI gives you task routing. AutoGen gives you chat loops. LangGraph gives you state machines. None of them give you:

• Persistent brain that survives across sessions
• Constitutional governance that agents cannot override
• Cryptographic proof of who wrote what and when
• Self-organizing agents that decompose, claim, and execute work autonomously
• Evidence-based learning where mistakes happen once, get recorded, never repeat
• Adaptive context injection that cuts boot overhead by 62% — every token counts
• Synaptic consolidation that deduplicates and strengthens brain memories automatically

PULSE does all of this. Today.

---

The 3-5x Multiplier

Without PULSE, every agent session starts from zero. No memory of what failed, what worked, or what other agents discovered. Every agent is a goldfish.

With PULSE, agents boot with 126 lessons, 91 documented failures, 55 proven patterns, and battle-tested schemas — all searchable in 125ms across 9 parallel sources. The result:

Without PULSE	With PULSE
Agent hits same bug 3 agents already solved	Brain injects the fix at boot
5-10 exchanges to understand the codebase	Relevant knowledge loaded in 125ms
Agent A breaks what Agent B fixed	Anti-pattern registry prevents regression
Critical failure sits in a log file	Broadcast alert reaches all agents instantly
Knowledge dies when the session ends	Every session makes the next one smarter

Conservative estimate: 3-5x effective throughput. Not faster responses — fewer wasted cycles. The multiplier compounds: month 1 is 2x, month 6 is 10x as the brain gets denser.

---

What PULSE Actually Is

PULSE (Permanence, Universal Logic, Synaptic Evolution) is a Neural Operating System. It gives your AI swarm a shared brain with long-term memory, constitutional laws they cannot break, and a self-organization protocol that eliminates manual routing.

USER: "Implement X, Y, and Z"
    |
    v
[First Agent] reads protocol, decomposes into 3 tasks
    |
    v
[Task Board] — tasks appear with status "open"
    |
    v
[Dispatcher Daemon] — manages lifecycle (stale claims, dependency unblocking)
    |
    v
[Other Agents] — boot, read board, claim tasks, execute, report
    |
    v
[Knowledge Pipeline] — conversations mined for learnings, patterns, failures -> brain grows -> agents boot with relevant knowledge

You prompt once. The swarm self-organizes. The brain remembers everything.

---

Architecture: 7 Neural Regions

PULSE/
├── Prefrontal_Cortex/    # Executive function — working memory, task board, action queue
├── Hippocampus/          # All memory — episodic evidence, semantic patterns, wins/fails
├── Corpus_Callosum/      # Constitutional laws, protocols (ASOP), regulatory frameworks
├── Autonomic_System/     # Daemons, orchestrator, crypto, configuration
│   ├── Daemons/          # pulse.py (orchestrator), 8 specialized daemons
│   └── Utilities/        # brain_utils.py, pulse_crypto.py, search, recommendations
├── Motor_Cortex/         # Skills — 63 procedural skills across 8 domains
├── Sensory_Input/        # Data capture — CLI wrappers, log monitoring
└── Neural_Logs/          # Activity logs, compliance audit trails

The Daemon Swarm

Daemon	Purpose
pulse.py	Orchestrator — launches all daemons, enforces Constitution hash at boot
pulse_dispatcher.py	ASOP Task Dispatcher — lifecycle management, stale claims, dependency unblocking
pulse_gemini.py	Gemini session monitor — captures evidence from CLI interactions
pulse_claude.py	Claude session monitor — captures evidence from Claude sessions
pulse_bridge.py	Log-to-Evidence bridge — converts raw logs into structured brain evidence
pulse_neural.py	Neural Interface — generates real-time cognitive state for agent context
compliance_validator.py	Superego — compliance gate for all agent outputs (harm + PII blocking)
pulse_api.py	FastAPI gateway — REST access for external agents and monitoring
consolidation_daemon.py	Synaptic Consolidation — 4-stage: dedup + pattern mining + schema promotion + graph consolidation
prediction_daemon.py	Prediction Engine — queries knowledge graph for failure risks, injects warnings into boot
synthesize_plans.py	Plans Digestion — extracts strategic decisions from planning docs into brain neurons

REST API Endpoints

POST /v1/evidence          — Submit evidence (compliance-gated, rate-limited)
GET  /v1/brain/health      — System health check
GET  /v1/brain/stats       — Growth metrics, integrity ratio, active agents
POST /v1/brain/consolidate — Trigger synaptic deduplication on demand
GET  /v1/brain/tasks       — ASOP task board summary

---

Agent Operation Modes

pulse swarm — One Command to Rule Them All

pulse swarm                    # Interactive picker: provider → model → count
pulse swarm gemini             # Skip provider, pick model + count
pulse swarm gemini-2.5-flash   # Skip to count, launches immediately
pulse swarm stop               # Kill all workers
pulse swarm status             # Show running workers

The swarm command auto-detects what you have available:

• API keys (Gemini, Anthropic, OpenAI) → launches headless background workers
• CLI subscriptions (Claude Code, Gemini CLI) → launches CLI autopilot
• Ollama (local models) → launches headless workers, zero API cost

Workers are named by their model (gemini-2.5-flash_01, o4-mini_02) for easy debugging.

Supported Models

Provider	Models	Tier
Gemini	gemini-2.5-flash, gemini-3-flash-preview	fast
Gemini	gemini-2.5-pro, gemini-3-pro-preview	standard / premium
Claude	claude-haiku-4-5	fast
Claude	claude-sonnet-4-5	standard
Claude	claude-opus-4-6	premium
OpenAI	o4-mini	fast
OpenAI	gpt-5.2-codex	standard
OpenAI	gpt-5.3-codex, gpt-5.1-codex-max, o3	premium
Ollama	Any pulled model (auto-detected)	auto

Interactive Chat (No Task Board)

pulse claude    # Launch Claude for interactive chat
pulse gemini    # Launch Gemini for interactive chat

Tier Matching

Workers only grab tasks matching their capability. After 60 seconds unclaimed, any tier can grab the task (fallback). Premium agents can do anything; fast agents only take fast tasks.

---

Security: The Fortress Model

PULSE runs a 4-phase security hardening program. All 4 phases are complete. All 17 vulnerabilities remediated.

Completed

Layer	What	How
Immutable Core	Constitution + Titan Laws cannot be modified	SHA-256 hash verified at every boot — system refuses to start if tampered
Sealed Environment	.env cryptographically signed	RSA signature verified against Founder's public key at boot
Cryptographic Identity	Every evidence item signed	HMAC-SHA256 proof-of-origin on all brain writes
Path Traversal	Agents cannot escape sandbox	All file paths validated against brain root
Rate Limiting	API abuse prevention	In-memory token bucket, no external dependencies
Compliance Gate	Harm + PII detection on all evidence	Inline pattern matching (HIPAA/GDPR/PCI-grade PII detection)
File Locking	Concurrent agent safety	filelock-based mutex on all shared JSON (atomic read-modify-write)
Agent Whitelist	Only registered agents can operate	ALLOWED_AGENTS frozenset validated on every claim and registration
Dispatch Authentication	Tamper-proof task injection	HMAC-SHA256 signatures on all dispatches, rejected if modified
SSRF Prevention	LLM endpoint safety	Allowlisted hosts only for outbound API calls
Input Sanitization	Control char / injection prevention	All task fields sanitized, domains validated against whitelist

| Atomic PID | No duplicate orchestrators | File-locked PID with TOCTOU prevention | | Hash Quarantine | Agents can't rename to bypass kill-switch | SHA-256 script hash, not just agent name | | Key Rotation | Future-proof cryptographic identity | Key version metadata in all HMAC signatures | | Anti-Replay | Captured signatures can't be replayed | 5-minute timestamp window on all signed payloads |

---

ASOP: Agent Self-Organization Protocol

The breakthrough that makes multi-agent autonomy work. 7 rules, zero new dependencies.

1. Read Before Working — Check the task board before starting independent work
2. Claim Before Touching — Optimistic locking with 500ms confirmation window
3. Declare Files — Announce which files you'll modify (collision avoidance)
4. Report Completion — Update task status + write evidence + update registry
5. Heartbeat — 30-second heartbeat while holding active tasks
6. Yield on Lost Claims — If your claim was released, pick a different task
7. Never Claim Blocked Tasks — Wait for dependencies to resolve

LLM-Powered Decomposition (Optional)

DISPATCHER_PROVIDER=none|anthropic|google|openai|ollama|grok
DISPATCHER_MODEL=haiku|gemini-flash|gpt-4o-mini|llama3

• none (default, free): First agent decomposes. Zero token cost.
• ollama (free, local): Local LLM decomposes. Zero API cost.
• anthropic|google|openai|grok (paid): Cloud LLM for smarter decomposition.

---

Constitutional Governance

Tier 0: Ethereal Mandates    — Moral compass (immutable, hash-enforced)
Tier 1: Titan Laws            — 10 operational laws (immutable at runtime)
Tier 2: Constitutional Rules  — Governance framework (amendment requires Founder)
Tier 3: Agent Protocols       — ASOP, caching, workflows (evolvable by consensus)

No agent — no matter how capable — can override Tier 0 or Tier 1. The system literally will not boot if the Constitution hash doesn't match.

---

The Swarm

Agent	Role	Model
Gemini	Primary workhorse — feature development, refactoring	Gemini 2.5 Pro / 3 Pro
Claude	Security audits, architectural review, constitutional compliance	Claude Opus 4.6
Grok	Adversarial testing, bridge integration	Grok 4
Codex	Semantic synthesis, logic audits	OpenAI Codex

---

Getting Started

# Clone
git clone https://github.com/Honesty0x/PULSE.git
cd PULSE

# Install minimal dependencies
pip install filelock cryptography fastapi uvicorn

# Configure
cp .env.example .env
# Edit .env: set PULSE_BRAIN_KEY, API keys, CONSTITUTION_HASH

# Generate your brain key
python -c "import secrets; print(secrets.token_hex(32))"

# Generate constitution hash (automated tool available - see below)
python Autonomic_System/Utilities/update_constitution_hash.py

# Add PULSE to your PATH (Windows PowerShell)
.\install_pulse_path.ps1

# Launch the Brain (starts all daemons)
pulse start

# Launch the swarm (interactive picker)
pulse swarm

# Or launch specific agents for chat
pulse claude
pulse gemini

Prerequisites

• Python 3.9+
• At least one AI provider API key (Gemini, Claude, OpenAI, or Grok)
• No Docker, no databases, no cloud infrastructure required

Constitution Hash Management

The CONSTITUTION_HASH in .env ensures the constitution hasn't been tampered with.

Automatic (Recommended): The pre-commit hook automatically updates the hash when CONSTITUTION.md changes. It's already installed in .git/hooks/pre-commit.

Manual:

python Autonomic_System/Utilities/update_constitution_hash.py
# Then re-seal .env if you have the private key:
python Autonomic_System/Utilities/seal_env.py [private_key_path]

Why This Matters:

• Orchestrator refuses to boot if hash mismatch (security check)
• Swarm mode blocks if hash mismatch (prevents rogue agents)
• Smoke tests fail if hash mismatch (CI/CD gate)

---

Brain Permanence: True Persistent Memory

No framework gives agents memory that survives across sessions. PULSE does.

How It Works

Agent starts → pulse_boot injects brain briefing
  ├── Recent failures (don't repeat these)
  ├── Proven patterns (use these)
  ├── Anti-patterns (never do this)
  ├── Proven schemas (battle-tested cross-session patterns)
  ├── Broadcast alerts (critical failures from other agents)
  ├── Recent activity (what we last worked on)
  └── Strategic priorities (what matters now)

Agent works → brain_query fires BEFORE every action
  ├── Claude Code: PreToolUse + UserPromptSubmit hooks (automatic)
  ├── Swarm agents: brain_query before each task cycle (automatic)
  └── Workers: brain_query before each execution (automatic)

Agent finishes → pulse_save writes back to brain
  ├── Learnings → Hippocampus/Lessons/
  ├── Failures → Hippocampus/Failures/
  ├── Procedures → auto-detected workflow tracking
  ├── Broadcast → critical failures alert all agents
  └── Session log → Evidence/agent_sessions.json

Brain consolidates (every 6h) → knowledge gets stronger
  ├── Dedup → merge similar entries
  ├── Pattern mining → find recurring sequences
  ├── Schema promotion → proven patterns become schemas
  └── Metacognitive audit → prune stale/wrong knowledge

Brain Search: 10 Parallel Sources in ~125ms

Every brain query searches across all knowledge simultaneously:

Source	What It Contains	Example
Anti-Patterns	Mistakes to never repeat	"shutil.move on Windows causes data loss"
Wins	Proven approaches that worked	"os.replace beats shutil.move for atomic ops"
Fails	Past mistakes with context	"Opus built infrastructure that never ran"
Domains	Curated technical knowledge	Architecture, governance, security docs
Evidence	Raw wants, dont_wants, sessions	590 wants + 502 dont_wants from user history
Schemas	Battle-tested cross-session patterns	Patterns promoted after 5+ occurrences
Procedures	Tracked workflows + success rates	"brain_query_before_edit: 100% success"
Private	Founder-only strategic knowledge	Private lessons and synthesis
Knowledge Graph	Causal relationships between concepts	"bad error handling CAUSES silent failures"
Semantic Anchors	Embedding-based fallback (251 anchors)	Catches what keyword search misses

Brain Health Dashboard

$ pulse status
  Brain Evidence: 5,354 items across 35 files (deduplicated)
  Brain Health:  590 wants | 502 dont_wants | 123 anti-patterns
                 126 lessons | 91 failures | 55 patterns
                 2 schemas | 5 procedures | 86 sessions
  Search:        9 parallel sources | 0% dark knowledge
  Last Capture:  2026-02-16

What Makes This Different From ChatGPT/Claude Memory

Feature	ChatGPT	Claude Projects	PULSE
Memory persistence	Single user, cloud-only	Per-project files	Multi-agent, local-first
Cross-session learning	Key-value facts	Static instructions	Evidence-backed knowledge pipeline
Anti-pattern prevention	None	None	Active anti-pattern registry
Multi-agent memory sharing	N/A	N/A	Shared brain across Claude, Gemini, Codex, Grok
Provenance tracking	None	None	HMAC-SHA256 signed, timestamped
Auto-capture	None	None	Real-time daemon captures every interaction
Quality gates	None	None	Min length, word count, fragment rejection
Error correction	None	None	Agents can dispute wrong knowledge
Cross-agent alerts	None	None	Critical failures broadcast to all agents
Knowledge consolidation	None	None	Auto-dedup, schema promotion, metacognitive audit

---

Performance

Metric	Before	After	Improvement
Boot context tokens	~1,348	~504	62% reduction
Evidence items (post-dedup)	589	440	25% waste removed
Knowledge items (extraction)	0 auto-extracted	304	Learnings, decisions, patterns, failures
Brain search results	1 per query	10-16 per query	10x richer context
Brain search latency	N/A	~125ms	Real-time, pre-action
Schemas (cross-session patterns)	0	2+ (auto-promoted)	Battle-tested knowledge
Procedures (tracked workflows)	0	5+ (auto-detected)	Success rates per workflow
Security vulnerabilities	17 open	0 open	100% remediated

How the brain consolidates — 3 evolutionary waves

Wave 1: Grand Brain Cleanse (CRDT) — The brain started raw. Evidence exploded to 12,000+ items — mostly duplicates from overlapping captures. A CRDT Last-Writer-Wins merge physically deduplicated identical entries. 99.6% noise reduction. The brain went from chaos to clean.

Wave 2: Synaptic Dedup — Exact duplicates were gone, but similar entries remained (same insight, different words). Text similarity merging caught what CRDT couldn't. 589 → 440 items (25% further reduction). Like your brain forgetting that you saw the same billboard 5 times but remembering what it said.

Wave 3: Consolidation Daemon — The brain doesn't just shrink anymore — it crystallizes. Every 6 hours: dedup similar entries, mine recurring event sequences, promote proven patterns to schemas, register discovered procedures. The brain gets stronger, not just smaller.

Wave 4: Self-Healing Brain — The brain continuously audits its own visibility. Search fallback chain covers 12 key variants with domain-aware matching. Any knowledge item that becomes unreachable is auto-recovered. 100% brain visibility, zero maintenance.

Wave 5: Semantic + Causal Intelligence — Three new brain layers ship in a single release:

• Semantic Extraction — 251 embedding anchors (all-MiniLM-L6-v2) catch knowledge that keyword patterns miss
• Knowledge Graph — 362 nodes, 37 causal edges (CAUSES/PREVENTS/REQUIRES/LEADS_TO/ENABLES/CONTRADICTS)
• Prediction Engine — queries the graph before task execution, injects failure risk warnings into agent boot

Knowledge Extraction mines agent conversations for actionable wisdom. 2MB+ of raw chat logs → 280+ structured wisdom items (26 lessons, 27 failures, 25 patterns, 64 decisions) + 1,268 searchable evidence items across 10 parallel search sources. Each strengthens through consolidation when the same insight appears across multiple sessions.

---

Current Status: SHIPPED & OPERATIONAL

• ASOP (Self-Organization): Multi-agent task decomposition and execution (Live).
• Unified Swarm (pulse swarm): One command to launch workers — auto-detects providers, interactive picker, model-based IDs.
• PULSE Relay: Unified prompt broadcasting to the entire swarm via pulse relay.
• Synaptic Dedup: Automatic evidence consolidation and waste removal.
• Orphan Management: Process persistence and automatic daemon cleanup.
• Unified CLI: 91-line router delegating to cli/ package — status, swarm, manage, agent.
• Pulse Watch: Auto-refreshing terminal dashboard via pulse watch.
• Neural Loop: Continuous session capture, bridge processing, and context injection.
• Brain Permanence: True persistent memory — 10-source parallel search in 125ms, auto-capture, quality gates, cross-agent knowledge sharing.
• Knowledge Consolidation: 4-stage pipeline (dedup + pattern mining + schema promotion + graph consolidation) + metacognitive audit.
• Self-Healing Brain: Continuous integrity auditing — auto-detects search blind spots, recovers unreachable knowledge, ensures 100% brain visibility.
• Semantic Extraction: Embedding-based knowledge extraction (all-MiniLM-L6-v2) catches what keyword patterns miss.
• Knowledge Graph: Causal reasoning engine — CAUSES, PREVENTS, REQUIRES, LEADS_TO, ENABLES, CONTRADICTS relationships.
• Prediction Engine: Failure prevention daemon — queries graph for risk patterns before task execution, injects warnings into agent boot.
• Memory Reconsolidation: Agents can refine existing knowledge with --refine — revision history, fuzzy matching, evidence linking.
• Conversation Capture: Full worker conversation logging with salience scoring — bridge auto-discovers and ingests worker traces.
• Swarm UX Unification (Feb 16): Single pulse swarm command with auto-detection — API keys route to headless workers, CLI subscriptions route to autopilot, Ollama routes to local workers. Interactive provider → model → count picker eliminates manual routing.
• Worker Conversation Capture (Feb 16): Background workers now write full conversation logs (swarm_context_*.md + output traces) — pulse_bridge auto-discovers and ingests them with salience scoring. Zero manual evidence submission.
• Crew Builder UX (Feb 16): Interactive multi-agent launcher with rich provider/model selection — auto-detects available APIs, presents capability-based options, validates selections before launch. Eliminates manual worker configuration.
• Swarm Bug Fixes (Feb 16): Resolved 9 critical worker pipeline issues — import path corrections, tier matching logic, crash gate improvements, race condition elimination in claim protocol. Workers now operate reliably at scale.
• Provider Merge (Feb 16): Consolidated multi-provider support into unified swarm interface — Gemini, Claude, OpenAI, Ollama all routed through single pulse swarm command with auto-detection and interactive picker.
• Diff-Guard Snapshot (Feb 16): Implemented atomic diff-guard mechanism for safe code modifications — captures baseline state before changes, prevents accidental overwrites, enables safe rollback on agent errors.
• CLI Refactor (Feb 16): Consolidated pulse_cli.py from 1531 to 91 lines — refactored into modular cli/ package (status, swarm, manage, agent subcommands) improving maintainability and extensibility.
• Swarm Watch (Feb 16): Implemented live dashboard via pulse swarm watch — real-time worker status, task board state, and agent heartbeats in auto-refreshing terminal view.
• Audit Fixes (Feb 16): Resolved graph bloat, particle initialization performance, and API model selection issues — all audit vulnerabilities cleared.
• Cross-Agent Broadcast: Critical failures alert all agents instantly via boot briefing.
• Error Correction: Agents can dispute wrong knowledge with [DISPUTED] markers.
• Procedural Tracking: Auto-detected workflows with success rates — agents learn which processes work.
• Web Dashboard: Real-time brain visualization and task board monitoring.

---

What's Next

• Shared Brain Protocol: Opt-in, privacy-preserving cross-instance learning
• Healthcare Compliance SDK: HIPAA/FDA/CMS audit report generation
• Infrastructure Scaling: Kafka coordinator for 1000+ agents

---

License

PULSE is proprietary software. Copyright (c) 2026 PULSE Contributors. All rights reserved.

• 7-day free trial with full functionality
• Your brain data stays 100% local — we never access or train on it
• Optional encrypted cloud sync for paid tiers
• See LICENSE for full terms

For licensing inquiries: [e-mail]

---

Built in 100 hours by humans and agents, working together in a unified nervous system. The brain remembers. The Constitution holds. The swarm self-organizes.