ledger

Ledger

Give your AI a memory it can never forget — and you can never lose.

Role-shaped, graph-queryable, 100% local. The memory layer behind the Quorum AI C-Suite — and a standalone knowledge graph for any Claude, MCP, or local-LLM workflow.

100% recall is the design requirement. We store your exact words. We never summarize. We never paraphrase. Your data physically cannot leave your machine — there is no cloud endpoint to call.

🎬 60-second demo: coming in the next release — watch a Claude session recall a 6-month-old decision verbatim, in 80ms, from a graph that has never touched the network. (Track the demo build →)

---

Before / After

You ask Claude: "Why did we pick Stripe over Lemon Squeezy back in May?"

Generic Claude (no Ledger):

I don't have memory of our previous conversations. Could you share what you discussed? Generally, the trade-offs between Stripe and Lemon Squeezy include: Stripe gives more control and better APIs but requires you to handle merchant-of-record duties; Lemon Squeezy handles MoR but takes a higher fee. Without your specific context...

Generic answer. You're the one with the memory. The AI just rephrases what you already know.

Claude + Ledger:

Pulling from cfo/decisions/2026-05-25-stripe-vs-lemon-squeezy:

"Reversed Lemon Squeezy → Stripe processor decision today. Infinidatum LLC is registered MoR, so the Lemon Squeezy MoR benefit is moot. Stripe gives better webhook ergonomics for the license-JWT issuance path. Cost delta: ~1.1% favor Stripe on volumes above $50K/mo."

The reversal happened 2026-05-25 in the launch-session batch. The CFO weighed in on the cost delta; the CIO weighed in on the webhook ergonomics. Logged in the same drawer as the marketplace-pricing decision.

Your exact words, returned verbatim, with the timestamp, the people involved, and the surrounding decisions linked.

---

Why it wins in 2026

What	Why it matters	Where it lives
Verbatim by design	We never summarize, paraphrase, or lossy-compress user data. We store your exact words and return them. If you said it, we kept it.	mempalace/searcher.py
Local-first by architecture	All extraction, chunking, embedding, retrieval, and graph traversal happens on your machine. External LLMs are BYOK, opt-in, never default, never silent fallback.	CLAUDE.md design principles
96.6% R@5 on LongMemEval (real, sourced)	Beats the published baselines on the standard long-conversation recall benchmark — using zero LLM calls at retrieval time. Pure index + graph + AAAK symbolic compression.	benchmarks/README.md
Graph-queryable, role-shaped	Entities, relationships, and temporal validity (valid_from / valid_to) are first-class. With sakthi init --sabha, the 9 Quorum role wings (CFO, CMO, CSO, CIO, CAIO, CXO, CHRO, CLC, CEO) come pre-bootstrapped — so memory naturally segments by domain.	mempalace/knowledge_graph.py
32 MCP tools	Full read/write/graph-traversal surface exposed to any MCP client (Claude Code, Claude Desktop, any MCP-aware host). Search, knowledge-graph queries, tunnel traversal, diary, sittham corpus ingest — all callable from the model.	MCP tool list
Pairs with Quorum	The Quorum council's 9 C-suite skills route to a role, then query the matching Ledger wing. The two repos ship as a marketplace pair, MIT-licensed, designed to compose.	Quorum

---

One-command install (full stack)

Ledger standalone — for any Claude / MCP / local-LLM workflow:

# 1. Install
uv tool install ledger        # or: pip install ledger

# 2. Bootstrap with the Quorum 9-wing preset (recommended even if you skip Quorum)
sakthi init --sabha ~/sakthi

# 3. Register with Claude Code as an MCP server
claude mcp add sakthi -- sakthi-mcp --palace ~/sakthi

Ledger + the Quorum council together — the full local-first AI C-Suite stack: