UPII — your memory, on your machine

Local-first memory for AI. Ingest your notes and documents, ask questions, get answers with citations — and no byte of your corpus ever leaves your device.

$ upii ask "What did we agree with ICEYE on resolution and latency?"

  Answer: You agreed on 0.5 m resolution imagery with a 6-hour delivery
  SLA, with latency relaxed to 12 hours during the pilot phase.

  Sources:
    [1] meetings/2026-03-14_iceye_sync.md   (chunk a3f2c9)
    [2] contracts/iceye_pilot_terms.md      (chunk 77d01b)

Status: early preview. The core capture, retrieval and ambient stack works and is tested, but interfaces are still moving.

Known gaps — worth knowing before you judge the claims above:

• The relational retrieval signal is off by default. Ingestion now populates a local knowledge graph (people/orgs/projects) and upii knowledge --graph renders it, but on the retrieval eval, letting the relational signal re-rank is not yet net-positive (it can boost a chunk that merely mentions a query entity over a better semantic match), so its fusion weight defaults to 0. Retrieval today is effectively semantic + (recency-weighted) temporal; raise the signal per query with upii ask --w-relational. Tuning it to help is ongoing.

Why UPII

Most RAG and "second brain" tools upload your corpus to someone else's cloud and rent your own memory back to you. UPII inverts that:

• 100% local. Embeddings (sentence-transformers MiniLM), vector store (LanceDB), metadata (SQLite), and reasoning (Ollama) all run on-device. A remote LLM (Gemini) is optional and swappable — never required.
• Every answer is cited. Answers are attributed to source chunks by content-addressed ID. Retrieval is verifiable, not just plausible.
• Deterministic, reproducible memory. Chunk IDs are a pure function of content + config. Re-ingesting the same corpus reproduces 100% of chunk hashes — validated by tests and benchmarks. Citations stay stable across re-indexes; re-embedding is cache-safe.
• More than cosine similarity. The Context Rehydrator is built to fuse three signals — semantic (vectors), temporal ("what was I working on last week"), and relational (a local knowledge graph of people, projects, orgs, populated on ingest) — into one ranked context window, with per-signal contributions visible via upii ask --debug. Measured, not asserted: Recall@10 = 0.958 on a committed labelled set (reproduce it). The relational signal is live but weight-0 by default until it's tuned to help (see Status); the score above is from semantic + temporal.
• Consent-gated ambient capture. A filesystem watcher keeps memory current, but sources are explicitly enabled and ambient events pass through an approval inbox before entering durable memory.
• Graceful degradation. No GPU, no API key, no internet? Retrieval still works and the system still answers.

Quickstart

Requires Python 3.10–3.12.

pip install upii            # or: pipx install upii

upii doctor                 # verify the local stack (db, vectors, model, disk)
upii ingest ./notes --recursive
upii search "satellite latency"
upii ask "What did we agree with ICEYE on resolution and latency?"

First run downloads the embedding model (all-MiniLM-L6-v2, ~90 MB) — the only download UPII ever needs. For local reasoning, install Ollama and ollama pull llama3.2; without it, UPII falls back to retrieval + a deterministic mock, so it always responds.

From source

git clone https://github.com/<you>/upii && cd upii
python3 -m venv venv && source venv/bin/activate   # Windows: venv\Scripts\activate
pip install -e ".[dev]"
pytest tests/ -q

Usage

upii ingest ./notes --recursive   # build memory from a directory (idempotent)
upii search "satellite latency"   # semantic + attributed retrieval
upii ask "…"                      # cited answer over your corpus
upii write "follow-up" --target email   # synthesize in your voice
upii watch ./inbox                # ambient capture (consent-gated)
upii inbox                        # approve/reject ambient captures
upii knowledge --graph            # render the local knowledge graph to HTML
upii sources list                 # manage enabled sources
upii metrics show                 # observability
upii doctor                       # health check

Run upii --help for the full command set.

How it works

graph TD
    User[User / Ambient Sources] -->|ingest| IngestionEngine
    User -->|ask / search| Rehydrator

    subgraph "Everything runs locally"
        IngestionEngine -->|hash + dedup| MetadataStore[(SQLite — source of truth)]
        IngestionEngine -->|deterministic chunking| Chunker
        Chunker -->|local embeddings| Embedder
        Embedder -->|store vectors| VectorStore[(LanceDB)]

        Rehydrator -->|semantic| VectorStore
        Rehydrator -->|temporal + relational| MetadataStore
        Rehydrator -->|fused context| LLM[Ollama local / optional remote]
        LLM -->|attributed answer| User
    end

Every write path — ingest, watch-approve, demo seed — flows through a single deterministic pipeline (src/upii/ingestion/pipeline.py):

• Identity is content-addressed. doc_id = hash(content), chunk_id = hash(chunk_text, config). No random UUIDs, no wall-clock, no path dependence. Same input ⇒ same memory state.
• Dedup — re-ingesting unchanged bytes is a no-op.
• Edit — a changed file re-chunks only what changed; unchanged chunk hashes stay stable; the prior version (chunks + vectors + metadata) is purged.
• Delete — removal cleans chunks, vectors, and metadata together.

Reproduce it yourself:

bash scripts/demo/repro_demo.sh                          # ingest → re-ingest → identical state
python scripts/bench/scale_check.py --docs 500 --paras 60   # scale + hash-reproducibility report

How UPII compares

Roadmap

• Tune the relational signal to net-positive — it re-ranks on knowledge-graph entity overlap but currently isn't a clear win on the eval, so it ships weight-0; make it help before enabling by default
• Batched vector writes — remove the per-document LanceDB append that caps ingestion throughput
• MCP server — expose upii_search / upii_ask as local MCP tools so Claude, Cursor, and other agents can use your private, cited memory (scope)
• Mail + calendar connectors (consent-gated)
• Answer verification / abstention
• Signed installers for macOS, Windows, Linux
• upii forget — user-facing verified deletion

Contributing

Issues and PRs welcome. Start with:

pip install -e ".[dev]" && pytest tests/ -q

Hard constraints for any change: local-first (no required cloud dependency), deterministic ingestion (same input ⇒ same chunk hashes), and consent-gating for any ambient capture path. See tests/test_chunk_determinism.py for the invariants that must hold.

License

MIT © 2026 Maddy