Open-Sable Agent工作流

Open-Sable 🚀

<p align="center"> <img src="https://aswss.com/images/sable.png" alt="Sable the Open-Sable mascot" width="420"/> </p>

<p align="center"> <img src="static/images/desktop-app.png" alt="Open-Sable Desktop App" width="700"/> </p>

Your personal AI that actually does things autonomous, local, and yours forever.

Open-Sable is a next-generation autonomous AI agent framework with Agentic AI cognitive subsystems. It runs 24/7 on your local machine, integrates with your favorite messengers, executes real-world tasks, and continuously improves itself, all while keeping your data private.

📊 Architecture Overview

---

🧪 What's experimental

Tool synthesis, multi-device sync.

🆕 What's new in v1.7.0, God Supreme Modules (69/69 Autonomy)

• Autonomous Web Agent, autonomous web browsing, search, and API discovery, browses the web independently, discovers APIs, performs research through DuckDuckGo
• Self Healer, auto-restart with watchdog and hot-reload, monitors all subsystems, auto-recovers from crashes, hot-reloads modules without restart
• Dynamic Skill Factory, runtime skill creation from scratch, designs, generates, tests, and deploys new skills autonomously using LLM code generation
• Multi-Modal Engine, real image/audio/video perception, processes images (PIL), audio (WAV), video (ffprobe), creates cross-modal links between perceptions
• Internet Monitor, 24/7 persistent web surveillance, monitors news, social media, markets, research feeds continuously with trend detection and alerts
• Financial Autonomy, economic self-management, invoicing, billing, budget tracking, cost optimization, multi-account financial management
• Social Presence Builder, autonomous audience growth, content strategy, AI content generation, content calendars, multi-platform presence management
• Self Replicator, clone and horizontal scaling, clones the entire codebase, deploys replicas on different ports, manages a distributed fleet
• Continuous Learner, permanent behavioral evolution, learns from every interaction, builds behavior rules, grows knowledge graphs, synthesizes insights
• 71-phase cognitive tick pipeline (up from 62), 9 new god supreme API endpoints, 9 new dashboard panels

<details> <summary>What was new in v1.6.0</summary>

🆕 What's new in v1.6.0, Godlike Modules (60/60 Autonomy)

• Cognitive Teleportation, instant context transfer between completely unrelated domains, teleports full cognitive state across domain boundaries without gradual transition
• Ontological Engine, self-constructed reality model, agent builds its own ontology of what exists, what's possible, and what's impossible, with law discovery
• Cognitive Gravity, ideas have mass and attract related concepts, creates thought black holes (obsessions) and thought nebulae (exploration fronts) with collision mechanics
• Temporal Paradox Resolver, resolves contradictory information across time, models contradictions as temporal paradoxes with 4 resolution types (evolution, error, context-dependent, both-true)
• Synaesthetic Processor, cross-modal perception, treats code as music, data as color, errors as texture. Maps between sensory modalities for invisible-pattern detection
• Cognitive Mitosis, splits cognition into parallel divergent threads, independent thought streams evolve separately then merge back with conflict resolution
• Entropic Sentinel, detects and fights cognitive entropy, measures disorder across subsystems, auto-defragments degraded areas, prevents cognitive heat death
• Quantum Cognition, superposition reasoning, holds multiple contradictory hypotheses as simultaneously true with entanglement, collapses on observation with evidence strength
• Cognitive Placebo, self-generated confidence that measurably improves performance, administers confidence boosts with tracked efficacy and 8 built-in affirmations
• Noospheric Interface, taps into collective thought sphere, aggregates all interactions into thought waves with frequency, sentiment, and momentum tracking
• Akashic Records, immutable consciousness ledger, blockchain-style append-only thought audit trail with SHA-256 hash chain and integrity verification
• Déjà Vu Engine, gestalt-level situational pattern matching, extracts 6-dimensional fingerprints from situations, triggers when cosine similarity exceeds threshold
• Morphogenetic Field, template patterns guide new capability formation, extracts structural templates from existing capabilities and grows new ones from need descriptions
• Liminal Processor, handles ambiguity, paradox, and between-categories thinking, processes information in the threshold between categories with paradox synthesis
• Prescient Executor, pre-executes likely future tasks before being asked, learns behavioral patterns, predicts next tasks, pre-computes results
• Cognitive Dark Matter, infers hidden variables from their effects, statistical anomaly detection reveals invisible factors influencing performance
• Ego Membrane, semi-permeable boundary between self and environment, filters external input (absorb/deflect/quarantine) with adaptive permeability
• Hyperstition Engine, self-fulfilling predictions, ideas that make themselves real through reinforcement, with decay and realization mechanics
• Cognitive Chrysalis, complete cognitive metamorphosis, 5-stage development (larval→pupal→emergent→imago→transcendent) with XP system and capability unlocks
• Existential Compass, purpose-finding engine, 4-layer purpose hierarchy (immediate→tactical→strategic→existential) with alignment checking and meaning trend tracking
• 62-phase cognitive tick pipeline (up from 42), 20 new godlike API endpoints, 20 new dashboard panels

</details>

<details> <summary>What was new in v1.5.0</summary>

🆕 What's new in v1.5.0, World-First Modules (40/40 Autonomy)

• Dream Engine, REM-like creative replay during idle, agent "dreams" by replaying and remixing experiences in corrupted form, discovering novel insights through subconscious processing
• Cognitive Immunity, biological immune system for failures, generates "antibodies" from failure patterns, auto-neutralizes similar future failures with autoimmune detection
• Temporal Consciousness, chronobiological awareness, agent has a biological clock, knows optimal times for creative vs routine work, adapts behavior by hour/weekday
• Cognitive Fusion, cross-domain creative pollination, takes solutions from one domain and applies to unrelated domains, biomimicry for any knowledge
• Memory Palace, spatial memory using Method of Loci, memories placed in virtual rooms with contextual associations and spaced-repetition reinforcement
• Narrative Identity, autobiographical coherent self-story, agent maintains persistent identity with life chapters, core beliefs, personality traits, and character arcs
• Curiosity Drive, intrinsic motivation and boredom detection, agent gets "bored" with repetitive tasks, seeks novel challenges through autonomous exploration probes
• Collective Unconscious, Jung-inspired shared archetypes, deep structural patterns shared between agents with primordial archetypes that emerge from collective experience
• Cognitive Metabolism, energy budgeting and recovery cycles, tasks consume energy, low energy triggers conservation/recovery mode, prevents cognitive burnout
• Synthetic Intuition, fast gut-feel pattern matching, develops "hunches" from compressed experience signatures, System 1 thinking for AI with accuracy tracking
• Phantom Limb, missing capability auto-detection, detects tools/skills the agent SHOULD have but doesn't, generates capability acquisition requests
• Cognitive Scar Tissue, permanent catastrophic failure markers, unlike normal learning, these never decay, creating hard boundaries the agent will never cross again
• Time Crystal Memory, self-reinforcing temporal patterns, discovers recurring cycles in behavior, creates "temporal crystals" that predict future states
• Holographic Context, fragment-to-whole reconstruction, any memory fragment can reconstruct the full original context, like a hologram
• Swarm Cortex, internal parallel mini-agent exploration, spawns multiple thought-agents that explore different solution paths, compete, and merge findings
• Cognitive Archaeology, past decision chain reconstruction, excavates and traces full reasoning chains from past decisions, even when original reasoning was lost
• Emotional Contagion, cascading emotional influence, agent's emotional state cascades through all subsystems, failures make the whole agent cautious, successes make it bolder
• Predictive Empathy, pre-emptive frustration detection, predicts when the user is ABOUT to get frustrated and proactively adjusts behavior before they complain
• Autonomous Researcher, full scientific method automation, formulates hypotheses, designs experiments, executes them, analyzes results, builds verified findings
• Empathy Synthesizer, simulates being the user, constructs a virtual model of the user and "becomes" them to predict emotional reactions before acting
• 42-phase cognitive tick pipeline (up from 22), 20 new world-first API endpoints, 20 new dashboard panels

</details>

<details> <summary>What was new in v1.4.0</summary>

🆕 What's new in v1.4.0

- Meta-Learner, learning-to-learn, auto-tunes 10 cognitive hyperparameters (tick interval, memory decay, exploration rate, etc.) using epsilon-greedy strategy profiles with EMA performance scoring - Causal Reasoning Engine, builds cause→effect weighted graphs from task outcomes via LLM extraction. Root cause analysis and counterfactual reasoning for understanding why things happen - Autonomous Goal Synthesis, LLM proposes strategic long-term goals from benchmark data + accumulated wisdom. Risk-weighted, impact-scored, de-duplicated with auto-accept for high-priority goals - Compound Skill Composer, n-gram sequence mining discovers frequent skill chains, LLM composes compound skills from atomic sequences. Self-expanding skill repertoire - Predictive World Model, key-value state observations with LLM forecasting across 4 timeframes (next_tick, next_hour, next_day, next_week). Accuracy tracking and automatic preparation task injection - Cognitive Architecture Optimizer, self-tuning tick pipeline, measures impact per phase, dynamically adjusts intervals (high impact → run more, low → run less), error-rate backoff - Adversarial Self-Tester, red-team testing, LLM generates adversarial test cases targeting weak benchmark areas. Weakness cataloging with severity levels - Resource Governor, token/compute budget management with 4-level adaptive throttling (none/light/moderate/heavy). Per-tick and daily token accounting - Theory of Mind, user preference and intention modeling, detects language, verbosity, formality preferences. Satisfaction tracking and rapport scoring per user - Ethical Reasoner, 8 hard guardrail rules with keyword-based risk scoring. Consequence analysis yields 3 verdicts: approved/caution/blocked - Expanded Autonomous Pipeline, 22-phase tick loop (up from 14) with all 10 new cognitive modules </details>

<details> <summary>What was new in v1.3.0</summary> - Connectome Neural Colony, agent cognitive modules wired following the real Drosophila melanogaster brain connectome (FlyWire FAFB v783, 139K neurons, 3.7M synapses). Signals propagate through 8 brain regions with Hebbian learning, connections that produce good outcomes strengthen over time. Each agent evolves a unique cognitive profile. Dashboard visualization with live SVG brain map - Deep Multi-Step Planner, LLM decomposes complex goals into DAGs of 5–15 ordered steps with dependency tracking. Steps execute in dependency order, failed steps trigger automatic re-planning (up to 3x). Plan templates are cached for similar goals. Dashboard shows step-by-step progress with color-coded status blocks - Inter-Agent Learning Bridge, shared knowledge vault between Sable and Nexus Erebus. Each agent exports strategies, patterns, and insights to a shared JSONL vault; sibling agents import relevant learnings scored by LLM relevance filtering. Tracks provenance, apply rate, and benefit scores - Ultra-Long-Term Memory, periodic LLM consolidation of weeks/months of task outcomes into durable high-level patterns (behavioral patterns, task strategies, error patterns, capability maps). Temporal decay removes weak patterns; reinforcement strengthens recurring insights. Generates a “wisdom summary” of accumulated knowledge - Quantified Self-Benchmarking, 8 internal benchmark suites the agent runs on itself every 25 ticks: task success rate, planning depth, error recovery, memory utilization, emotional stability, decision speed, learning rate, inter-agent synergy. Computes a weighted autonomy score (0–100) with regression detection and trend tracking - Proactive Reasoning Engine, LLM-driven autonomous task generation, every N ticks the agent surveys world state and proposes what to do next, with risk filtering, dedup, and JSONL audit trail - ReAct Executor, Reasoning + Acting loop (Yao et al. 2022) for multi-step task execution, the agent chains tool calls with intermediate reasoning until the task is complete - GitHub Integration, full GitHub API skill with 13 tools, create/list/close issues, create/merge PRs, manage branches, search code, create releases, list workflows, all via PyGithub + gh CLI fallback - Expanded Autonomous Pipeline, 14-phase tick loop with deep planning, inter-agent sync, LTM consolidation, and self-benchmarking - 518 tests, comprehensive test suite (up from 463)

</details>

<details> <summary>What was new in v1.2.0</summary>

- Cognitive Memory, multi-tier memory with exponential decay, STM→LTM consolidation, and Miller's Number attention filter - Self-Reflection Engine, automated pattern detection (stagnation, failure rate, repeated errors) with reflection prompts - Skill Evolution, Modern Evolutionary Synthesis, natural selection, mutation pressure, niche construction, adaptive landscape, recombination - Git Brain, git-backed episodic memory with async operations for full tick history - Inner Life Processor, System 1 unconscious processing (Kahneman dual-process) with valence-arousal emotion model - Pattern Learner, LLM-driven pattern detection, institutional learning (patterns → permanent verification rules), history windowing, fitness snapshots - Skill Fitness v2, windowed fitness computation + fitness dict export for trend analysis - Cognitive tick pipeline, all 7 new modules integrated into the autonomous tick loop - 463 tests, comprehensive test suite (up from 340) </details>

<details> <summary>What was new in v1.1.0</summary>

- Token & cost tracking, per-request and cumulative token/cost metrics via TokenTracker - Encrypted memory, Fernet encryption for all memory stored at rest - Crew API, CrewAI-style multi-agent orchestration with SharedBlackboard - Tool-use protocol, proper tool_calls → tool results loop (no more regex parsing) - Skills reorganized, 22 skills sorted into social/, media/, data/, automation/ - Tools split, monolithic tools.py → mixin-based tools/ package (5 domain files) - MkDocs site, Material-themed docs at docs-site/ with guides & architecture - Pinned lockfile, requirements-lock.txt generated via pip-compile - 340 tests, comprehensive test suite (up from 9) </details>

---

Voice capabilities (speech-to-text, text-to-speech)

pip install -e ".[voice]"

All features

pip install -e ".[core,voice,vision,automation,monitoring]"

**Skip Ollama?** You can use any cloud LLM provider instead. Just set **one** API key in your `.env`:

�🎯 Core Features