ERMI Command Center

Externalized Recursive Memory Infrastructure, built as a local-first MVP.

This implementation ingests ChatGPT-style conversations.json exports and ChatLasso SSI Markdown, preserves raw source data, creates an Obsidian-compatible Markdown vault, indexes conversations/messages/chunks/entities/SSI metadata in SQLite, generates embeddings, supports hybrid search, surfaces regression flags, and exports a lightweight graph.

Project

ERMI is maintained as one of Jusstin's KnightBot projects.

Canonical project aliases live in docs/ALIASES.md.

Quick Start

After cloning on Windows, double-click:

START-HERE.cmd

That runs the one-click installer, diagnostics, desktop shortcut setup, and app launch.

To build a Windows setup wizard executable, see docs/SIGNED_WINDOWS_INSTALLER.md.

For Android/Termux mobile use and sideload APK builds, see docs/ANDROID_TERMUX.md.

Manual CLI flow:

python -m ermi init
python -m ermi setup --chatgpt-source C:\path\to\conversations.json --chatlasso-source C:\path\to\Obsidian\10_Data_Harvest\11_SSI_Raw --run
python -m ermi diagnostics
python -m ermi chatgpt-titles C:\path\to\conversations.json
python -m ermi ingest C:\path\to\conversations.json
python -m ermi export-chatgpt-csv C:\path\to\conversations.json
python -m ermi mine-chatgpt-code C:\path\to\conversations.json
python -m ermi chatgpt-activity C:\path\to\conversations.json
python -m ermi export-chatgpt-obsidian C:\path\to\conversations.json
python -m ermi import-chatlasso C:\path\to\Obsidian\10_Data_Harvest\11_SSI_Raw
python -m ermi watch-chatlasso C:\path\to\Obsidian\10_Data_Harvest\11_SSI_Raw --once
python -m ermi search "recursive memory architecture"
python -m ermi flags
python -m ermi timeline
python -m ermi entities
python -m ermi graph
python -m ermi backup

Command Center UI

Run the local API and frontend in two terminals:

npm run api
npm run dev:ui

Then open:

http://127.0.0.1:5173

The UI provides a local command center for first-run setup, health diagnostics, ingesting exports, watched ChatLasso folders, hybrid recall, entity inspection, regression flags, import review, concept timeline, graph export, backups, archive counts, and operations logs.

The ingest panel supports both raw ChatGPT exports and ChatLasso SSI Markdown output, making ChatLasso the capture/synthesis layer and ERMI the durable recall/index layer. ChatGPT imports follow the export's current_node path, so regenerated/abandoned branches stay out of the main archive. ChatLasso can also POST SSI Markdown directly to ERMI at http://127.0.0.1:8765/api/import/chatlasso-payload. ERMI preserves ChatLasso mode, archetype, status, hash_beacon, loss_report, audit state, regression flags, and domain nodes as structured metadata.

Windows Desktop Installer

After cloning the repo on Windows, double-click:

install\Install-ERMI.cmd

The installer checks core prerequisites, installs ERMI dependencies, initializes the local archive, and creates two Desktop shortcuts:

• ERMI Command Center
• Update ERMI Command Center

Update Workflow

git pull
npm install
python -m pip install -e ".[dev,ml]"
python -m ermi --root archive migrate
npm run api
npm run dev:ui

See docs/MVP_HANDOFF.md, docs/DEE_TORCH_POST_CLONE.md, docs/SIGNED_WINDOWS_INSTALLER.md, docs/ANDROID_TERMUX.md, docs/GETTING_STARTED.md, docs/INSTALLER.md, docs/OPERATIONS.md, and CHANGELOG.md for ongoing project notes.

See docs/CHATLASSO_INTEGRATION.md for the joined ChatLasso -> ERMI workflow.

By default ERMI writes to:

archive/
  setup.json
  raw/
  vault/
    conversations/
  ermi.sqlite3
  graph.json
  watchers.json
  backups/

Sample smoke-test data lives under:

sample_data/

Optional ML Embeddings

Without optional packages, ERMI uses a deterministic local hashing embedder so the system works offline immediately. For stronger semantic retrieval, install:

pip install -e .[ml]

Then re-run ingestion so chunk embeddings are regenerated with sentence-transformers.

Design Notes

• Raw source files are copied, never overwritten.
• Markdown and database rows are derived artifacts and can be rebuilt.
• ChatGPT mapping graphs are reconstructed from the final current_node path when available.
• Chunk embeddings are stored in SQLite for a simple local MVP.
• ChatLasso SSI payloads are imported as first-class memory artifacts with schema-versioned metadata.
• Hybrid search uses local embeddings plus SQLite FTS5, so exact phrases and semantic matches both matter.
• Flagged imports enter the review queue; clean imports are accepted automatically.
• Entity extraction is deliberately conservative and local. It can be replaced later with an LLM or NLP pipeline.
• The graph export is JSON for portability; NetworkX, Neo4j, or RDF can be layered on later.