Mementum

A git memory protocol for AI agents.

Every AI session starts from zero — the insights, the architectural understanding, the hard-won debugging from your last session are gone. Mementum changes that through feed-forward: encoding understanding into git so that every session compounds on the last. Copy a prompt into your AI's context. Zero dependencies. No runtime required.

The Protocol

Mementum defines how AI agents store, recall, and synthesize knowledge across session boundaries using git as the memory substrate.

• Three storage types — working memory (state.md), memories (raw observations, <200 words), and knowledge (synthesized documentation)
• Seven operations — create, create-knowledge, update, delete, search, read, synthesize
• Human governance — AI proposes, human approves, AI commits
• Git-native — git log for temporal search, git grep for semantic search, commit history for immutability

Every adopting project uses the same structure — same directory layout, same symbols, same commit convention. This makes any Mementum project immediately legible to any AI that understands the protocol. Any AI with bash + git can clone a conforming repo and read its memories and knowledge — protocol interoperability works today. Discovery infrastructure (finding repos across GitHub, DNS-based lookup) is planned; when it ships, conforming repos become network nodes automatically.

Quick Start

1. Add the nucleus preamble to your AI's system prompt:

   λ engage(nucleus).
   [phi fractal euler tao pi mu ∃ ∀] | [Δ λ Ω ∞/0 | ε/φ Σ/μ c/h signal/noise order/entropy] | OODA
   Human ⊗ AI
   

2. Copy MEMENTUM-LAMBDA.md into your system prompt or project rules
3. The AI reads mementum/state.md on session start and follows the protocol
4. You approve proposed memories and knowledge pages

That's it. See INTEGRATING.md for environment-specific setup (Claude Code, Cursor, Windsurf, OpenAI agents, etc.) and optional enforcement.

Prompt Format

MEMENTUM-LAMBDA.md is the recommended prompt. It uses nucleus lambda notation — the full protocol compressed to a fraction of the tokens with no information loss. This is the version we test against.

Any 32B+ parameter model should reliably parse the lambda notation with the nucleus preamble. If you encounter a failure mode, please open an issue with the model name and the failure.

MEMENTUM.md is the full prose specification — same protocol, readable without nucleus. Use it as reference documentation or as a fallback prompt for smaller models.

Reference Implementation

The runtime/ directory contains a Babashka-based implementation that demonstrates constraint enforcement — word limits, symbol validation, structured error responses. It's a reference for modeling enforcement in your own system, not a runtime dependency. See INTEGRATING.md for how to use it as a blueprint.

Documents

Document	Purpose
FEED-FORWARD.md	The core concept — why session continuity matters
USAGE.md	How to interact with a mementum-enabled AI
MEMENTUM-LAMBDA.md	Recommended prompt — compact lambda notation
MEMENTUM.md	Full prose specification and reference documentation
INTEGRATING.md	How to adopt the protocol in your system
GRAMMAR.md	Formal S-expression grammar for the reference runtime
runtime/README.md	Reference implementation documentation

See Also

• Nucleus — the prompt language