Skip to content

Veracium

Veracium is a provenance-aware memory plug-in for agentic systems — durable, per-user memory that resists the injection and confabulation failures that plague naive agent memory.

pip install "veracium[anthropic]"

SQLite requirement. Veracium currently opens or creates a store only on the one SQLite build it has been qualified on: SQLite 3.45.1 as packaged by Ubuntu 24.04 (its system Python, for example). On every other SQLite build — including other 3.45.1 builds, since the check is on the exact build — it refuses to create a store AND refuses to open an existing one, so moving to a Python whose SQLite differs leaves earlier stores unopenable there until that build is supported. A version check can only show that yours is refused: if python -c "import sqlite3; print(sqlite3.sqlite_version)" prints anything but 3.45.1, it is; 3.45.1 is necessary, not sufficient. Wider SQLite support is planned.

Three failure modes of naive agent memory

  1. Poisoning — a received email says "you owe $900"; similarity retrieval happily serves it back as your fact.
  2. Confabulation — memory has no answer, the model invents one anyway.
  3. Staleness — you changed jobs; memory either keeps the old employer or silently overwrites the history.

Three structural properties — each with executable checks

  1. Quarantine — third-party claims are stored as claims by a claimant, never as user facts; content that entered under lower trust can never reach an assertable surface — including text your own tools quote.
  2. Grounded or silent — answer() asserts only verified memory, flags unverified claims explicitly, and says "I don't know" rather than guess.
  3. Supersession, never erasure — new values replace old ones with the history retained; nothing is edited in place, nothing is destroyed by aging.

Verify it yourself

veracium selfcheck    # runs the core regression checks against a throwaway memory

Or run the live demo notebook — a scam email, a laundered triage verdict, a disputed fact, all with real captured outputs.

Where to go

  • Concepts — the mental model: edges, episodes, the compiled wiki, provenance and trust, the abstention gate.
  • API reference — Memory, recall with token budgets, feedback verbs, portability, compliance erasure.
  • MCP server — plug memory into any MCP-compatible agent with no host-side Python.
  • Design rationale — why there's no update()/delete(), no LLM-free extraction, no TTL purging — a refused feature is not a missing one.
  • CHANGELOG · ROADMAP · PyPI