v0.9.54 — approaching v1.0

Memory that cites its sources.

One Python library: save, search, ingest — every answer with its source.

In plain words

This page is for developers. An SDK is a piece of software you build into your own program.

Not a developer? Then Provenize Desktop is the product meant for you: the same functions, as an ordinary app.

From document to answer

This is the part for developers.

Guide figure pointing at the answer in the scene

Read the document in → cut into snippets → make it searchable → search → answer with its source.

Install

How it actually works today.

python -m venv venv
venv\Scripts\activate      # Windows
pip install -e .           # from the SDK folder

pip install -e .[ollama]     # local model
pip install -e .[mcp]        # link to AI tools
pip install -e .[postgresql] # multiple users

The SDK is not in a public package index yet, so pip install provenize does not work — we will only put that command here once it is true. Design partners get access to the code.

Three commands
from ai_memory import AIMemory

mem = AIMemory("./vault")
mem.ingest("reports/report-2025.pdf")

hits = mem.search("TLS misconfiguration")
print(hits[0].text)
print(hits[0].citation)  # report-2025.pdf · p. 41 · 0.93
What it can do

Everything below ships today, labeled honestly.

Core memory & retrieval

  • People, matters and how they connectBuilttyped entities and relations
  • Database for multiple usersBuiltmultiple users, each sees only their own data
  • Search on words and on meaningBuiltsearch on meaning and on words
  • Tasks that run on their ownBuilttasks that run on their own, with limits
  • The search index stays on your computerBuiltthe index never leaves your computer
  • Models are interchangeableBuiltmodels are swappable
  • AIMemory APIBuiltsave · search · get · delete · ingest
  • Database in a single fileBuiltreliable storage and fast search
  • Reading documents inBuiltPDFs, tables and scanned documents
  • Fixed fields per sourceBuiltsource, page and how well it matches
  • Checks what your computer can handleBuiltadapts to your hardware

Citations & provenance

  • Signed check number over the whole chainBuiltof the whole chain, checkable by outsiders
  • Signed evidence filesBuiltexports and erasure receipts
  • Every result carries its sourceBuiltfile, page and how well it matches — no source, no result
  • Refusal inside the answer pathIn developmentthe SDK does not sit inside the model; the link to AI tools (MCP) now exists in its first form
  • A log where each line is attached to the lastBuiltlines can only be added; tampering shows
  • Recheck it yourself (verify())Builtshows whether anything was edited, inserted or deleted

Privacy toolkit

  • Test for leaks yourselfBuiltrunnable by you or your client
  • People’s rightsBuiltexport, erasure, restriction
  • Removing personal detailsBuiltbefore anything goes to an external model

Access & multi-user

  • Access enforced in the databaseBuiltenforced in the database
  • Every automated action loggedBuiltevery automated action goes into the log
  • Retention-hold roleBuiltblocks erasure while retention applies
  • Rights per roleBuiltengagement-bound, enforced in the database

Engineering posture

  • The boundaries are written downBuiltshared with design partners
  • Continuous security checksBuiltcode, keys and known holes checked automatically
  • 2,500+ testsBuilt~90% coverage — our own test run; the report goes to design partners

Build toward v1.0 with us.

Become a design partner