Memory Security Skills Roadmap Install GitHub

Sovereign agent memory

Your AI should remember like a friend.

Vraksha is a local-first agent that restores your rules, project state, and personal context before every session. No amnesia. No borrowed memory. Just continuity on your machine.

Vraksha character portrait
memory 3 layers restored
trust score action gated
Quick Start
$ curl -fsSL https://raw.githubusercontent.com/vraksha/vraksha/main/install-linux.sh \
  | bash
Local store SQLite FTS5 + WAL
Security model multi-gate execution
Extensibility drop-in skill registry
Providers OpenAI, Anthropic, local, more

The thing most agents fake

Memory is the operating system.

Vraksha keeps durable knowledge in a local tri-store foundation so every conversation begins with your actual world loaded.

01

Procedural wiki

Rules, identities, hard boundaries, project-wide truths, and standards that should never drift.

02

Semantic recall

Past decisions and session context are indexed for retrieval instead of left behind in dead chats.

03

Relational metadata

Entities, projects, trust hints, and references prepare the path toward graph-backed reasoning.

Guarded autonomy

Give the agent hands. Keep the gates.

Vraksha is being built for machine operations with real scrutiny: sanitize inputs, verify intent, score actions, and keep execution isolated.

Gate 1

Input sanitization

Prompt injection and untrusted instructions are screened before they touch the reasoning loop.

Gate 2

Action verification

Shell commands, file writes, and external operations move through trust-scored checks.

Sandbox

Execution boundary

Docker isolation is the baseline today, with stronger sandbox targets on the roadmap.

Already useful

Not a demo skin. A growing agent system.

Async

Journal writer

Memory consolidation and logging stay away from the reasoning hot path.

Skills

Runtime registry

New experts load from folders with skill.py and SKILL.md.

Forensics

Slop detector

A specialized analyzer helps detect AI-generated code artifacts during review.

Models

Provider choice

Use hosted frontier models, OpenRouter, Bedrock, Groq, Ollama, and local endpoints.

Where it is going

From remembered work to private orchestration.

  1. Now Local memory foundation
  2. Next Gate logic inside the loop
  3. v1.0 Kuzu graph, MCP, Composio, Telegram
  4. Later A2A protocol and Guardian service