Persistent, fully offline local AI memory with a terminal UI.

Axoineo ingests your local documents, deduplicates them with SHA-256 hashing, embeds them into a persistent local vector database, and answers questions about them using a local LLM — all on your machine, with no network calls. The interface is a responsive terminal dashboard styled after the Claude Code CLI.

┌ ◆ AXOINEO ───────────────┬ ◆ MEMORY ─────────────────┐
│   mascot + version       │ chunks / docs / on-disk    │
│                          │ models · ollama status     │
├──────────────────────────┴────────────────────────────┤
│ conversation (Markdown answers, sources, system logs)  │
├─────────────────────────────────────────────────────────┤
│ ● ready          ❯ ask your memory anything, or /help   │
└─────────────────────────────────────────────────────────┘

• Ingest — SimpleDirectoryReader parses .pdf, .txt, .md, .docx, .xlsx. Files are chunked with a sentence splitter and embedded via Ollama.
• Deduplicate — a file is hashed before parsing (skip re-work on an unchanged file), and every chunk is hashed on whitespace-normalised text (skip boilerplate shared across documents). A JSON manifest beside the store records file hashes so a file whose chunks are all already known is still remembered and never re-parsed.
• Retrieve — the query engine pulls the top-k chunks from ChromaDB and the local LLM answers strictly from that context, citing source files.
• Persist — everything lives in ./axoineo_memory/. Delete it to reset, copy it to move your memory to another machine.

• Python 3.10+
• Ollama running locally, with two models pulled:

ollama pull qwen2.5:1.5b
ollama pull nomic-embed-text

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python main.py --doctor    # verify dependencies + Ollama connectivity
python main.py             # launch the TUI

--doctor runs on a bare interpreter and reports exactly which dependency or model is missing, so run it first if anything fails to start.

Inside the TUI, type a question to search memory, or use a slash command:

Keys: ctrl+l clear · ctrl+b toggle dashboard · f1 help · ctrl+c quit.

Ingestion and generation run on background threads (@work(thread=True)), so a long vectorisation or a slow LLM response never freezes the UI; a spinner in the status line shows progress.

Every setting has an environment-variable override — no code edits needed:

AXOINEO_DB_PATH=/media/usb/brain \
AXOINEO_LLM_MODEL=qwen2.5:3b \
AXOINEO_OLLAMA_URL=http://localhost:11434 \
python main.py

See axoineo/config.py for the full list (chunk size, top-k, timeouts, max file size, …).

python main.py              # launch the TUI
python main.py --doctor     # dependency + Ollama preflight
python main.py --wipe       # delete the memory bank (asks first)
python main.py --db PATH    # use an alternate memory bank
python main.py --version

The engine's deduplication, manifest, retrieval and error handling are covered by an offline suite that fakes the Ollama/Chroma/LlamaIndex stack — no models, network, or heavy dependencies required:

python -m unittest discover -s tests -v

axoineo/
├── config.py   # paths, model names, timeouts, tunables (all env-overridable)
├── engine.py   # LlamaIndex + ChromaDB, SHA-256 dedup, manifest, stats
└── tui.py      # Textual dashboard + threaded workers
main.py         # entry point (TUI + --doctor/--wipe)
tests/          # offline test suite