"Building AI that doesn't just predict—it perceives, remembers, regulates, plans, and acts."
Shiva is a command-line cognitive human like-agent which acts as an integrated cognitive model rather than just a simple wrapper around an AI model. In a nutshell it is a general intellligence. Directly rivals the General Intelligence Model of China.
Rather than building another conversational chatbot, Shiva aims to build a Cognitive General Intelligence model capable of continuously planning, reasoning, executing code, remembering, and adapting directly inside developer workspaces.
Most workspace AI tools today are purely reactive, the agents are robotic, relying on probabilistic ways of coding and generating the necessary language. As a founder and CTO, would you like your coding agent or email drafting agent to be robotic or like a person who does all of it, with no data being sent to external servers, the person is a software. Yes correct !! This is why we are building shiva.ai
Intelligence should not begin with a prompt and end with a response. Real-world autonomous General Intelligence systems must continuously think, reason, remember previous executions, self-critique, and execute sandbox operations without manual prompt-chaining.
Shiva models workspace automation as a distributed Cognitive Operating System rather than a monolithic neural network. It decomposes cognition into specialized subsystems responsible for distinct aspects of intelligence, coordinating them through a shared runtime.
The language model is not the brain. It is a reasoning engine used by the cognitive system whenever probabilistic inference, planning, or code generation is required.
[ Dev Workspace / CLI ]
│
▼
Workspace Context Hook
(Grep / Directory Scans)
│
▼
Semantic Understanding & Memory
│
▼
Mothership Arbitration Loop
│
┌───────────────┴───────────────┐
▼ ▼
Cognitive Swarm Internal Regulation
│
▼
Chain of Thought
(Thought • Critique)
│
▼
Sandbox Execution
│
▼
Monte Carlo Assignment ─────── SQLite MemoryGraph
- Unified DTOs: Clean data tracking using a consolidated DTO model in
src/transferDTO.py. - Integrated Sandbox Execution: Reasoning is compiled into executable Python scripts, run in a secure local sandbox with rich standard libraries, and results are appended directly back to the decision chain.
- Autonomous Idle Processing: When the CLI is waiting for user prompts, Shiva automatically runs sleep and dream consolidation cycles in the background, suspending instantly on user input.
- Monte Carlo Credit Assignment: At the end of each reasoning episode, trajectory steps are evaluated and reinforced in a vector-less SQLite MemoryGraph.
Shiva columns generate multi-turn reflections, critiques, and decisions. If a column outputs a Python 3 block to read/write files or execute calculations, the Decoder executes it in the local sandbox under a WORKSPACE_DIR context, returning standard outputs back to the decision flow.
When idle, Shiva runs background sleep threads utilizing DreamGenerator and the consolidation/forgetting/identity models. As soon as the user enters a prompt, the dream thread immediately stops, commits memory states to SQLite, and executes the query.
Shiva resolves multi-step credit assignment using Monte Carlo estimation. Trajectory node paths and temporal edges are mathematically reinforced:
strength and association_strength inside the SQLite memory repository based on the final query confidence.
###4. Mothership and cells Remember "The Avengers" film, battle of new york, where an army of chitauri had come in. the two legged creatures were being launched by a big flying whale. shiva.ai functions the same way were it as a mothership which deployes cells for solving a problem. Each cell has its own strength.
- Memory: Custom vector-less SQLite
MemoryGraphwith topological activation spreading. - Regulation: Prefrontal
Mothershipcoordinator and physical pendulum stability regulator. - Appraisal: FT-Transformer-based emotional dynamics appraisal engine.
Shiva is a Cognitive Operating System that transforms existing developer workspaces into autonomous cognitive runtimes, continuously planning, executing, remembering, and adapting.
Shiva is an open-source research project. We welcome contributions in Deep Learning, Reinforcement Learning, Cognitive Science, and Software Automation.
Released under the Apache 2.0 License.
Intelligence is more than prediction. It is perception, regulation, memory, adaptation, and action.