EVM — Evolution Vector Memory

Related publication

The model-agnostic continuity architecture underlying EVM has been formalized separately as:

Interaction Dynamics Layer (IDL) 1.0: A Model-Agnostic Continuity Architecture for Long-Term AI Systems

DOI: https://doi.org/10.5281/zenodo.20974299

While IDL defines the model-agnostic continuity architecture, EVM provides a reference framework implementing this architecture through directed Evolution Vector objects, extraction methods, and persistent interaction-state management.

---

A deterministic interaction-state framework for long-term AI continuity

Author: Szabolcs Krehlik
ORCID: 0009-0003-8623-7876
License: CC BY-NC-ND 4.0
Contact: szabolcs.krehlik@gmail.com

This repository contains the reference implementation accompanying the Evolution Vector Memory (EVM) framework. It provides a local web application with persistent storage, directed Evolution Vector (EV) extraction, PEV/EEV/FEV tracking, memory recall, semantic indexing, diagnostics, and long-term interaction-state continuity.

EVM is not a semantic vector database and not merely an affect or sentiment layer. It is a bounded trajectory architecture: every interaction is represented as a directed state transition, and long-running interaction identity is maintained through reconstructable PEV/EEV/FEV dynamics.

Name clarification

The name Evolution Vector Memory replaces the earlier working name Emotion Vector Memory. The EVM abbreviation is preserved, but the full name is corrected to reflect the actual scope of the system: interaction continuity, bounded identity evolution, deterministic trajectory reconstruction, and stateful AI behavior across time. See NAME_CHANGE.md for the rationale.

The core interaction object is a directed EV time-slice:

(x1, y1, z1, g1, e1, w1) -> (x2, y2, z2, g2, e2, w2)

This repository is intended for research, inspection, testing, and demonstration. It is not presented as a production deployment.

What is included

• Local Flask web app for chat-style interaction
• SQLite-backed persistence under DATA/sqlite/
• Plain NDJSON log mirrors under DATA/logs/
• EV entry/exit extraction per interaction
• PEV / EEV / FEV state tracking
• Pre-response planning and bounded state control
• Memory recall and indexed retrieval
• CIS export and diagnostics panel
• Provider support for OpenAI, xAI, and LM Studio

Repository structure

app/
  app.py                Flask routes and orchestration
  db.py                 SQLite access layer
  evm_core.py           EVM state update logic
  evm_spec.py           normative prompts and spec text
  indexing.py           turn summary and keyword indexing
  memory.py             retrieval helpers
  openai_client.py      provider adapter and model calls

app/templates/
  index.html            local web UI

DATA/
  logs/                 NDJSON mirrors created at runtime
  sqlite/               SQLite database created at runtime

install.command         first-time setup
start.command           start with provider from .env
start_openai.command    convenience launcher
start_lmstudio.command  convenience launcher
start_selector.command  provider chooser
run.py                  local server entry point