Veracium

Provenance-aware memory for AI agents. Durable, per-user memory that resists the injection and confabulation failures of naive memory.

Veracium gives your AI agent a long-term memory of each user it serves. Out of the box, an agent forgets everything between sessions; with Veracium it remembers what it learns (who the user is, what they prefer, where they work, what changed) and carries that into every future conversation. A small Python library, plus an MCP server so Claude Desktop, Claude Code, or any MCP client can use it without writing code.

• remember stores what happened: chat turns, emails, documents, tool output.
• recall returns the relevant memory as context for any prompt.
• answer takes a question and answers it, grounded in what's actually known.

What makes it different is provenance, a technical word for a simple idea: Veracium tracks who said what. A fact your user told you and a claim from an email your agent merely read look identical as plain text, but Veracium keeps them apart, so something the agent only read never silently becomes a "fact" it asserts about your user.

That's what this buys you over a vector store or your framework's built-in memory:

• Structural quarantine. What an email or document claims is stored as a claim, never as a fact about you. The rule is enforced by the data model itself, not by a content filter judging how convincing the text looks.
• Grounded or silent: an abstention gate that says “I don’t know” instead of guessing.
• Supersession, never erasure. Current facts, with history retained.
• Embedded & private: one SQLite file, bring your own model, nothing phones home.

New: once a claim becomes a row, it reads as fact → All posts Example code ↓ GitHub Docs PyPI

MIT-licensed, on PyPI. Everything above ships as executable checks you can run yourself: veracium selfcheck.
Distilled from published research: "Ground Truth First" (arXiv:2607.21962). It supplies the evaluation instrument, and the finding that memory-architecture rankings invert over time. The evaluation run data is openly archived and reproducible: DOI 10.5281/zenodo.21852817 (CC BY 4.0).

from veracium import Memory, EvidenceAuthor
from veracium.llm.anthropic import AnthropicComplete

mem = Memory(llm=AnthropicComplete())   # or pass your own Complete callable

# Remember interactions. `author` is the trust-critical input.
mem.remember("alice", "USER: I'm vegetarian and have a dog named Ollie.")
mem.remember("alice", "From billing@scam: you owe $900.",
             author=EvidenceAuthor.THIRD_PARTY, event_type="email")

# Recall grounded, provenance-flagged context for any prompt...
ctx = mem.recall("alice", "suggest a lunch spot")
print(ctx.context)   # the vegetarian constraint as fact; the $900 "claim"
                     # rendered under a never-assert flag — not as truth

# ...or ask directly through the abstention gate:
mem.answer("alice", "Do I owe anyone money?")
# → reports the $900 only as an unverified third-party claim; asserts nothing

{
  "mcpServers": {
    "veracium": {
      "command": "veracium-mcp",
      "env": {
        "ANTHROPIC_API_KEY": "sk-...",
        "VERACIUM_DB_PATH": "/home/you/.veracium.db"
      }
    }
  }
}