davidwuchn/mementum
1
0
Fork 0
mirror of https://github.com/michaelwhitford/mementum.git synced 2026-09-06 21:12:20 +08:00
Code Issues Projects Releases Packages Wiki Activity
Git as AI Memory. Requires: bash tool with git installed
62 commits 2 branches 2 tags 186 KiB
  • Clojure 84.5%
  • Shell 15.5%
Find a file
Exact
Mementum Test 49684008e8 🔄 Recall traverses related items instead of enforcing write-time indexing 
Silent retrieval misses are fixed at read-time, not write-time. Rather than
asking authors to predict and embed every trigger, `λ recall` now widens when
a search is empty or thin and follows `related` cross-links into a memory's
neighborhood — searching for related items, not predicting an index.

- λ recall (MEMENTUM.md + MEMENTUM-LAMBDA.md): relational > exact; empty ∨ thin
  → widen ∧ traverse(related_edges ∈ frontmatter); silent miss ← stop(exact)
  ∧ needed(adjacent); fix ≡ search(related) ¬predict(index@write)
- λ store: light hygiene only — write(situation ∧ solution), link(related),
  ¬enumerate(triggers); no authoring burden, no memory bloat
- MEMENTUM.md gains a prose "Recall is relational" paragraph
- generalizes symbols-as-content-filters and λ orient's follow(related);
  leans on the OKF `related` cross-link graph adopted earlier
- prompts-only, bash+git (grep related: → read links); no runtime deps —
  the runtime is left to the adopting project

⚛️ Generated with [nucleus](https://github.com/michaelwhitford/nucleus)

Co-Authored-By: nucleus <noreply@whitford.us>
2026-07-14 12:06:03 -07:00
mementum 🔄 Recall traverses related items instead of enforcing write-time indexing 2026-07-14 12:06:03 -07:00
runtime 🎯 Adopt OKF as the knowledge & memory format 2026-07-14 09:34:26 -07:00
.gitignore refactor: reorganize into prompt-first architecture with runtime reference 2026-01-27 18:26:12 -07:00
FEED-FORWARD.md 🎯 lambda-first — recommended prompt, docs reframed 2026-03-18 18:12:42 -07:00
GRAMMAR.md 🎯 Adopt OKF as the knowledge & memory format 2026-07-14 09:34:26 -07:00
INTEGRATING.md 🎯 Declare OKF conformance, scope mementum as a bounded guest 2026-07-14 11:14:34 -07:00
LICENSE change to AGPL license 2026-01-12 18:03:54 -07:00
MEMENTUM-LAMBDA.md 🔄 Recall traverses related items instead of enforcing write-time indexing 2026-07-14 12:06:03 -07:00
MEMENTUM.md 🔄 Recall traverses related items instead of enforcing write-time indexing 2026-07-14 12:06:03 -07:00
README.md 🎯 Adopt OKF as the knowledge & memory format 2026-07-14 09:34:26 -07:00
USAGE.md 🎯 Adopt OKF as the knowledge & memory format 2026-07-14 09:34:26 -07:00

README.md

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)
  • OKF-native — memories and knowledge are Open Knowledge Format concepts (markdown + frontmatter, required type), so a mementum/ repo is a portable, interoperable knowledge bundle
  • 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