Skip to content
navywakuraPublic

About

ECHO-5 research: objects, causal memory, planning and robotics. Validation ongoing; S6 remains open.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

4 Commits

Folders and files

Repository files navigation

ECHO-5

ECHO-5: research in progress Terminal release: 4.5 beta S6 under investigation Deterministic core ECHO 4.0 on Hugging Face

Python NumPy Textual HTTPX llama.cpp MCP

ECHO-AI studies object and relation representations, causal memory, program composition and planning through interaction. It combines a local cognitive core with an optional language adviser. This repository contains the core, research modules, development tests, echo-dev, the echo-nexus terminal and the new ECHO-5 implementation.

ECHO-5 is under active research and has not passed its complete release criteria. Cognitive experiments have produced audited results; Gazebo/PX4 confirmation still has unresolved failures. This source publication is not a green release. S6 remains open and its next investigation is deferred until after ECHO-5.5.

The 11 October checkpoint records what is implemented, what was measured and what remains unresolved. Selected receipts preserve positive and negative results with provenance. The 10 October archive retains the earlier B24/ARC work.

ECHO-5 now includes typed object and relation programs, concept reuse, hierarchical search, bounded hypothesis testing, an adaptive counterexample curriculum, MiniGrid/BabyAI evaluation, a MissionAgent bridge and a local decision viewer. The curriculum is an empirical adversarial-learning experiment; it is not a trained generative adversarial network.

The separate echounity sandbox provides an isometric robot world, a first-person camera, an observed map and genuine recordings of its decisions. Its simulation results have their own scope and do not close ECHO-5.

The existing terminal package remains 4.5.0b1. The previous ECHO-4.0 artifact is available on Hugging Face. Installing either package does not install a validated ECHO-5 release.

Work on ECHO-5

Start with the ECHO-5 setup, closure checklist and module inventory. ECHO-5 has its own Python environment and optional sandbox dependencies. The commands below install the existing terminal distribution.

The viewer presents observations, hypotheses, predictions, actions, consequences and remaining budget from actual records. Its neural diagnostic is separate from the acting policy. No fabricated inner monologue, consciousness claim or guarantee of zero hallucinations is implied.

Start

git clone https://github.com/navywakura/echoai.git
cd echoai
python3 tools/source_setup.py verify
bash install.sh
echoai setup
echoai tui

Python 3.12+, Linux or macOS, no administrator access required. The installer verifies every distributed source file, creates a private virtual environment, installs the included TUI and verifies the installed source again. It keeps an installation log and preserves existing non-symlink executables. Private repositories require GitHub authentication to clone. Windows users can run the source with Python; the shell installer requires WSL or a compatible shell.

Downloadable packages and SHA-256 checksums are in Releases. A release includes the wheel, complete source archive, bundled terminal harness and release evidence.

Connect the neocortex

Inside echo-nexus:

/connect api https://provider.example/v1 MODEL API_KEY_ENV
/connect openrouter free @/absolute/path/key-file
/connect anthropic https://api.anthropic.com/v1 MODEL ANTHROPIC_API_KEY
/connect local /absolute/path/model.gguf
/connect opencode provider/model
/connections save my-neocortex
/connections default my-neocortex
/devtest list
/devtest core-test
/echo

API connectors support OpenAI-compatible Chat Completions and Anthropic Messages. OpenRouter uses HTTPS. Local GGUF uses a separately installed llama-server; models are not downloaded automatically. OpenCode is a separately installed CLI running in an ephemeral adviser workspace with tools denied. Saved profiles contain credential references, never key contents. MCP stdio servers can be connected and invoked explicitly with /mcp; they are not automatically controlled by the adviser.

Model text is a proposal, not a verified fact. Connecting a language adviser does not enable the NEXUS-0 core cortex. /echo reports recorded observations; /ask explicitly shares a summary with the adviser. No fabricated thought stream is presented.

Terminal preview

Actual echo-nexus help screen

Develop

echoai verify
echoai doctor
echoai setup list
echoai devrun core-test
echo-dev doctor
echo-dev scan
echo-dev run

The setup TUI lists bounded development commands and captures their output in logs. echo-dev also supports isolated checks and controlled repair loops; its full profile requires Linux user namespaces, Bubblewrap and the separately managed tools described in the development harness guide. The setup launcher does not execute sealed S6 exams. Developers may use --allow-modified for their own edited checkout; modified source is never reported as a pristine release.

core/        Integer representation, associative memory and plasticity
nexus0/      Perception, policy, world model and epistemic gate
echo4/       Cognitive agent, persistence and causal-memory modules
echo_arc/    ARC perception and action solvers
seed/        Program synthesis, memory and S6 research controllers
harness/     echo-dev, echo-nexus and integration protocols
tests/       Development and regression tests
docs/        Specifications, research reports and current roadmap
tools/       Source verification and terminal setup
ECHO-5/      New research runtime, local viewer and controlled evaluations
research-archive/  Preserved experiment receipts and source provenance

The former echoai/ folder is now the repository root. Its required mathematical helpers live in core/; no obsolete chatbot or website is shipped. The live laboratory workspace and frozen experiments remain separate. Raw experiment environments and sealed input data are excluded from distribution.

research-archive/ contains explicitly preserved research snapshots, separate from live lab/ directories. Their manifests identify original paths, hashes, deduplicated files and omitted external build binaries. Installing this source does not execute those scripts or resume their experiments.

The ECHO-5 publication adapts imports to this repository's core/ layout. Laboratory receipts identify the original frozen sources; their hashes are not hashes of this portable distribution. See distribution notes.

Read current status, the roadmap, distribution notes and the terminal guide. Historical reports preserve their original paths and hashes.

Licensing

See LICENSE and component notices. The terminal harness is MIT; historical ECHO-4.0 packaging has its own notices under release1/LICENSES/. A beta package does not change those terms. Developer: RxLabs.

About

ECHO-5 research: objects, causal memory, planning and robotics. Validation ongoing; S6 remains open.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages