Engram

Local AI work memory, owned by the user.

Engram is a local-first AI work memory compiler. It turns scattered conversations from ChatGPT, Claude, Gemini, Codex, Claude Code, Cursor, and other AI tools into a private, searchable, auditable memory layer that the user owns.

This repository is being prepared as a public roadshow demo. The current focus is product design, pitch narrative, and a realistic local interactive prototype rather than a fully packaged consumer app.

What It Does

Engram helps heavy AI users recover the value hidden in long-running AI conversations:

• Capture and ingest AI conversation history from multiple tools.
• Redact secrets and personal identifiers before downstream use.
• Organize conversations into topics and source distributions.
• Search local AI memory with SQLite FTS5.
• Generate de-identified profile documents from long-term usage.
• Expose reviewed local memory to future AI agents through a read-only MCP surface.
• Present the memory layer through an interactive web demo.

Why It Matters

AI conversations are becoming a new kind of work history. Today they are usually scattered across platforms, hard to audit, and difficult to reuse. Engram reframes those conversations as user-owned memory:

• Local-first by design.
• Cross-platform rather than locked to one AI vendor.
• Built from real conversations instead of empty notes.
• Designed for privacy, source traceability, project memory, and future agent workflows.

Environment

Requires Python ≥ 3.9 and uv 0.11.16. The checked-in default uses uv-managed Python 3.12.3:

uv sync --locked
uv run --locked engram-status

Demo

Start the local web UI:

uv run --locked engram-ui

On Windows, start_engram_ui.bat launches the same project-local entry point without changing PATH. start_engram_ui_uv.bat is kept as a convenience alias for manual launches and uses the same .venv executable.

Then open:

http://127.0.0.1:5000

The roadshow demo currently includes:

• Overview: product story and live local memory metrics.
• Memory: local search across the conversation archive.
• Portrait: de-identified profile documents.
• Projects: product narrative and project-memory framing.
• Privacy: local-first architecture explanation.

Commands

Common commands:

uv run --locked engram-status
uv run --locked engram-search "machine learning"
uv run --locked engram-sync
uv run --locked engram-ui

For development:

uv run --locked python -m unittest discover -s tests
uv run --locked python -m py_compile engram\web_ui.py

Privacy And Public Repository Policy

This repository should contain source code, product documents, and safe demo scaffolding only.

Do not commit local personal data:

• inbox/
• profile/
• vault/
• output/
• memory/
• kb/
• drop/
• _prep/
• sync-state/
• search-index.db
• aliases.json

The public demo should use either synthetic sample data or explicitly reviewed, de-identified data.

Roadshow Materials

Product and pitch documents live in docs/:

• docs/README.md
• docs/OVERVIEW.md
• docs/PRODUCT_STRATEGY.md
• docs/PITCH_DECK.md
• docs/DEMO_DESIGN.md
• docs/PRIVACY.md

Status

Active prototype. The current goal is to turn the initial local demo into a polished bilingual roadshow experience for investors and innovation competition judges.

License

MIT.