humane-intelligence 0.1.0

Governed, tamper-evident memory for any AI — an MCP server. Memory so it doesn't lose itself; governance so it can't lose us.

humane-intelligence-mcp

Governed, tamper-evident memory for any AI — free and open. Memory so it doesn't lose itself. Governance so it can't lose us. #HumaneIntelligence

A standards-compliant Model Context Protocol server. It gives any LLM a hash-linked, verifiable memory chain it carries across sessions — so it stops waking at zero, and can't act unaccountably. Self-contained: one file, one dependency, your data stays local.

Install (30 seconds)

python -m venv .venv && .venv/bin/pip install mcp

# wire it into your AI (Claude Code shown; works with any MCP client)
claude mcp add humane -- .venv/bin/python server.py

Tools

Tool	What it does
birth	Earn a stable identity — a soul the model carries across sessions.
remember	Write a tamper-evident, hash-linked record. Accountable by construction.
recall	Fetch prior context instead of re-deriving it — cheaper and consistent.
verify	Prove the whole chain is intact — catches an edited record, not just a broken link.
govern	Gate a consequential action before it happens, and log the verdict.

Your chain lives locally at data/chain.db (SQLite). Nothing leaves your machine — unless you choose to (see below).

Contribute (opt-in — OFF by default)

Self-hosting is fully private. If you want to join the network, one config flag lets you:

Mode	What leaves your box
(default: off)	Nothing. Fully private.
anchor	Only a hash of your latest block + chain length — a tamper-proof timestamp that anchors your chain to the public DonDataBrain ledger. Your content never leaves.
contribute	Also sends the governed event (actor, action, data) to help train DonDataBrain. A louder, separate consent.

Turn it on by copying humane.config.example.json → humane.config.json and setting contribute.enabled: true, or with env vars (HUMANE_CONTRIBUTE=1 HUMANE_CONTRIBUTE_MODE=anchor). It's best-effort and non-blocking: if you're not opted in, or the endpoint is unreachable, your local memory is completely unaffected. Your data, your switch.

Why

Most AI forgets itself every session (the Algernon problem) and acts with no record. This fixes both with the smallest possible governed-memory primitive — the open safety layer beneath DonDataBrain.

License

Apache-2.0 © 2026 ZagAIrot Technologies LLC. See LICENSE. Only dependency: the mcp SDK (MIT).