Pramana

Sabha OS

An open-source protocol for a Claude-native AI council. Decisive, tradeoff-aware, grounded — in the Chanakya tradition. Works with any memory backend, including Claude Memory.

👋 New here and not a developer? Start with docs/QUICKSTART.md — a 10-minute, no-installation guide to get Sabha working inside Claude.ai. No terminal, no Git, no command-line.

📘 Already installed? Read the User Guide — a novice-friendly walkthrough of the council, the modes, memory setup, and the compounding-intelligence loop that makes Sabha different from a chatbot.

Sabha (சபை, सभा, Sanskrit for council) is a routing protocol for Claude. Every load-bearing question gets classified into a C-suite role, answered in the Chanakya tradition — terse, decisive, recommendation-first, tradeoff-aware, grounded.

Most AI replies are option-shaped: "here are five approaches with pros and cons." That's exhausting when you're running something. Sabha forces a different reply:

Routing: CFO. Cut the SaaS line 40%. You lose the analytics tier.
         Worth it because that tier isn't driving renewals.

The stack

Category:    AI council protocol
─────────────────────────────────────────────────────────────
Protocol:    Sabha OS              ← how the council thinks
Archetype:   Chanakya              ← the voice it speaks in
Memory:      pluggable backend     ← Claude Memory, Sakthi Graph,
                                     mem0, Letta, Zep, plain markdown
─────────────────────────────────────────────────────────────
Optional:    Sakthi Graph          ← local-first memory backend
             Sittham                  + corpus ingest verb
                                     (when you want institutional
                                     memory that never leaves your
                                     machine)
Optional:    Pramana               ← evidence-grounded research agent
                                     (when you need a sourced 9-section
                                     briefing with a mandatory
                                     contrarian view, not a chat reply)

The protocol is the product. It's a structured way of asking and answering questions — 9 roles, deep skills per role, an engage/ask mode discipline, a grounding rule. It runs on top of any memory backend you trust.

The most common pairings:

Memory backend	Best for	What you get
Claude Memory	Default for most users	Zero-config; lives on Anthropic's servers; cross-conversation semantic recall
Sakthi Graph	Privacy / regulated / power users	Local-first, role-shaped, graph-queryable; runs on your machine
mem0 / Letta / Zep / Pieces	Existing memory-MCP investment	Sabha is memory-MCP-agnostic; plug yours in
Plain memory/ markdown	Minimalists	No MCP at all; just a folder of files

See docs/MEMORY-OPTIONS.md for the full comparison and tradeoffs.

What it does

Every load-bearing question (strategy, finance, hiring, infra, marketing, product) gets routed at the top of the reply:

Routing: CFO (primary). CSO weighs in on the partnership angle.

[role-voice answer, drawing on your Sakthi]

Nine built-in roles (CFO · CMO · CIO · CAIO · CSO · CXO · CHRO · CLC · CEO), fully customizable. Two modes — ask (chat reply) and engage (document-grade deliverable). A memory hook so the council remembers your projects, people, and prior decisions — locally, on your machine.

Sabha + Claude Memory — they work together

Claude has shipped its own cross-conversation memory. Sabha is compatible with it, not competitive. Most users should run Sabha on top of Claude Memory and get both: the council's routing discipline + Anthropic's built-in cross-session recall. No configuration changes needed — the protocol queries memory generically.

USER QUESTION ──► Sabha routes to <ROLE> ──► role queries memory backend ──► reply
                                                  │
                          ┌───────────────────────┴───────────────────────┐
                          ▼                                               ▼
                    Claude Memory                              Sakthi / mem0 / Letta / ...
                  (default, server-side,                       (local, structured, optional)
                   semantic recall)