Hierarchical Memory
4-tier memory system (L0-L4) modeled after human cognition. With the Ebbinghaus Forgetting Curve.
MeshCtx is a 17-brain-region emulated self-improving agent platform. It remembers everything, learns from every task, and orchestrates multiple AI agents — open-core: AGPLv3 framework + source-available brain.
4-tier memory system (L0-L4) modeled after human cognition. With the Ebbinghaus Forgetting Curve.
Meta-cognition loop after every task: evaluate → extract patterns → update knowledge graph → adjust behavior.
One intent. Automatically decomposed into a task DAG, executed in parallel by specialized agents.
Learns your patterns. Pre-loads relevant memories before you ask. Context assembly in under 50ms.
MCP protocol native. Community-driven ecosystem of tools and integrations.
AGPLv3 framework. Free forever. Core brain algorithms source-available. Commercial license available.
Full-brain emulation with 17 brain regions: PFC Working Memory, Visual Cortex, NAcc Reward Prediction, Brainstem Autonomic, Hippocampal Replay, Amygdala Salience, Default Mode Network, Thalamic Gate, Cerebellar Forward Model, Basal Ganglia selector, ACC conflict monitor, Insula interoception, Mirror Neuron ToM, STDP plasticity, IIT integration, LTP enhancement, and Gnostic Field intuition. No other AI agent has this.
Sharp-wave ripple memory consolidation during idle. Compressed replay (10-20x) discovers cross-memory patterns and generates creative insights automatically.
Background creative ideation during idle periods. Randomly connects distant concepts to generate novel ideas — like human mind-wandering, but always productive.
Safe Python/Bash/JS execution. Docker isolation + subprocess fallback. Streaming SSE output. /run command in Chat.
Search your entire codebase. 15+ languages symbol extraction. Auto-watch file changes. /context command for instant retrieval.
Sparse Distributed Memory. O(2^1000) address space — 10^296x more capacity than any other agent. Predictive pre-activation + 100:1 fractal compression.
Industry-first: AI agent that autonomously analyzes, optimizes, tests, and applies changes to its own source code. 7-stage pipeline with safety guard.
4-phase operation contract: Propose→Verify→Commit→Reject. Critical actions require explicit user approval before execution.
Learns your time/context patterns. Pre-computes answers before you ask. Idle-time processing for instant responses.
Multi-pass iterative refinement with convergence detection. Re-examines and improves responses until quality threshold is met.
Real-time performance monitoring. Auto-adjusts cache, batch size, timeouts. PID-controlled parameter optimization.
Yann LeCun's Joint Embedding Predictive Architecture. Latent-space world model that predicts outcomes without pixel-level reconstruction — energy-efficient reasoning.
Windows GUI automation via Win32 API + UIA. Click, type, screenshot, window management — all through natural language commands.
Learns your approval patterns over time. Auto-approves routine operations, escalates risky ones. 3-tier: YOLO / Smart / Manual.
Heuristic-based model selection: routes queries to the best-fit model based on task type, content length, and cost constraints.
Tracks git activity (commits, churn, PR throughput) and session metrics. Generates weekly development activity reports.
不是噱头。每个脑区解决一个实际痛点:
闲时自动回放记忆,10-20倍速压缩。就像你睡觉时大脑在整理白天经历一样,它在后台默默发现规律——第二天回答问题更准。
自动识别消息的紧迫度和情感。"紧急!服务器挂了"和"今天天气不错"会被完全不同的优先级处理。
不等你提问,后台持续联想远距离概念产生新想法。别的Agent是单向工具,它像主动思考的同事。
海量信息中自动聚焦关键信号,抑制无关干扰。像嘈杂咖啡馆里仍能听清对面说话。
执行动作前先在"脑内模拟"后果。危险操作自动预警,防止AI闯祸。
TD学习持续优化策略。成功操作变"肌肉记忆"——越常用越快。
实时监测预期vs实际差异。发现偏差自动修正——不需要你手动纠正。
不只是字面理解,而是推断你真正的意图和情绪。急躁时简洁高效,探索时创意建议。
持续监控资源使用和性能。内存泄漏、响应变慢在你发现之前自我诊断报警。
保持当前任务上下文,快速切换注意力。多任务处理不丢上下文,Planning + WorkingMemory协同。
Gabor滤波器组提取纹理/边缘/方向特征。多尺度视觉缓冲区,适合图像分析和模式识别。
TD学习预测奖励信号,"想要vs喜欢"双通道。动机强度动态调节,驱动智能体主动探索。
自主神经调节系统+网状激活。心跳/呼吸/觉醒度三稳态控制,ReticularActivation保证系统持续在线。
Spike-timing-dependent plasticity. Synaptic weights adapt based on causal spike patterns — neurons that fire together, wire together.
Integrated Information Theory Φ-value computation. Measures causal integration as a proxy for system-level consciousness.
Activity-dependent synaptic strengthening via NMDA receptor simulation. Calcium-mediated kinase cascades for memory consolidation.
Non-symbolic holistic pattern detection. Pre-conscious gnostic field processes entire input gestalts — instant recognition without sequential reasoning.
📊 Benchmark验证 v3.115.18:8模块满分100% · 综合63.5% · 29/42通过 · 4个新脑区已集成
🔒 你的数据,你做主。所有操作可审计、可拦截、可回滚。API Key加密存储。框架AGPLv3开源——代码全透明,永不锁定。
| Capability | MeshCtx | Hermes | OpenClaw | Claude Cowork | FreeEnergy Chat |
|---|---|---|---|---|---|
| 🧠 JEPA World Model LeCun Latent Prediction | ✅ v3.36 | ✗ | ✗ | ✗ | ✗ |
| 🖥️ Desktop Agent Windows GUI Automation | ✅ v3.37 | ✗ | ✗ | ✗ | ✗ |
| 🔐 Smart Permissions Learns When to Auto-Approve | ✅ v3.38 | ✗ | ✗ | ✗ | ✗ |
| 🧠 Smart Router Heuristic Model Selection | ✅ v3.39 | ✗ | ✗ | ✗ | ✗ |
| Hierarchical Memory | L0-L4 | ✗ | ✗ | ✗ | ✗ |
| Forgetting Curve | ✓ | ✗ | ✗ | ✗ | ✗ |
| Meta-Cognition | ✓ | ✗ | ✗ | ✗ | ✗ |
| Multi-Agent | ✓ DAG | ⚠ Limited | ✓ | ⚠ Limited | ✗ |
| Plugin Market | ✓ 5 Plugins | ✗ | ✗ | ✗ | ✗ |
| MCP Protocol | ✓ | ✓ | ✓ | ✓ | ✗ |
| Amygdala Salience | ✓ | ✗ | ✗ | ✗ | ✗ |
| Counterfactual Reasoning | ✓ | ✗ | ✗ | ✗ | ✗ |
| Default Mode Network | ✓ | ✗ | ✗ | ✗ | ✗ |
| Thalamic Attention Gate | ✓ | ✗ | ✗ | ✗ | ✗ |
| Theory of Mind | ✓ | ✗ | ✗ | ✗ | ✗ |
| Multi-Model Compare | ✓ 3-Model Compare | ✗ | ✗ | ✗ | ✗ |
| 🛡️ Safety Guard | 4-Phase Contract | ✗ | ✗ | ✗ | ✗ |
| 🎯 Iterative Reasoning | Multi-Pass Refine | ✗ | ✗ | ✗ | ✗ |
| 🔬 Causal Root Cause (Pearl) do-calculus causal inference | ✅ v2.77 | ✗ | ✗ | ✗ | ✗ |
| 🔒 Prompt Injection Shield | 10 Inj. Types | ✗ | ✗ | ✗ | ✗ |
| 🔍 Cross-Validation | Multi-Agent OK | ✗ | ✗ | ✗ | ✗ |
| 📋 Behavior Compliance | 6 Rules | ✗ | ✗ | ✗ | ✗ |
| 🧮 Smart Router | Heuristic Select | ✗ | ✗ | ✗ | ✗ |
| 🔄 Self-Updater | push→test→deploy | ⚠ Manual | ✗ | ✗ | ✗ |
| 💾 Backup Vault | Multi-Path Backup | ✗ | ✗ | ✗ | ✗ |
| 🎯 Goal Decomposer | 4 Templates+DAG | ✗ | ✗ | ✗ | ✗ |
| 📚 Error Learner (ALiFE) | Never Repeat | ✗ | ✗ | ✗ | ✗ |
| ⚡ Workflow Engine | 11-Step Pipeline | ✗ | ✗ | ✗ | ✗ |
| 🐝 Agent Swarm Manager-Worker Multi-Agent | ✅ v3.34 | ✗ | ✗ | ✗ | ✗ |
MeshCtx is an open-core multi-agent framework built on "Division of Labor + Full Transparency." It's not a black-box single agent — it's a mesh network of an Orchestrator, a Guardian, and specialized agents working under full human supervision.
Questions? Collaborations? Reach out anytime.
DMG package · ARM64 native · Homebrew
Source install · pip · venv · PEP 668
Community-driven · MCP native · GitHub hosted · One-click install
Google/Bing/DuckDuckGo search integration
v1.0 · 0 installsAI code review: bug detection/security/style
v1.0 · 0 installsLocal file browse/read/write/search
v1.0 · 0 installsFeishu bot: messages/notifications/group management
v1.0 · 0 installsData visualization: charts/dashboards/reports
v1.0 · 0 installs