Skip to content

What is Aevum?

Aevum is a Python library that gives AI agents a signed audit trail, consent-checked data access, and verifiable decision records — three problems that tend to surface together in production. The quickstart gets you to working code in ten minutes.

Consent enforcement. Data cannot be accessed without an active consent grant that specifies exactly who can access it, for what purpose, and for how long. Revoking consent takes effect at the next operation — no batch job, no delay.

A cryptographic audit trail. Every operation produces a signed, chained audit record. The record is tamper-evident: changing any past entry breaks the cryptographic chain and is immediately detectable on verification.

Human approval gates. Actions that require human sign-off cannot proceed without a signed approval event in the audit trail. If no one responds before the deadline, the action is blocked. Silence is a veto.

What it is not

Aevum is not a memory framework. It does not replace Mem0, LangChain, or LlamaIndex. It sits below those tools and makes whatever memory layer you are using consent-gated and auditable.

Aevum is not a SaaS product. You install it in your own infrastructure. Your data never leaves your environment.

Aevum is not an AI orchestration framework. It does not run agents or manage workflows. It governs what happens when an agent accesses data.

Aevum does not make your application compliant with any regulation. It provides technical controls — tamper-evident audit records, consent documentation, and human-review gates — designed to support compliance programs. Whether those controls satisfy a specific regulatory obligation depends on your deployment, configuration, jurisdiction, and surrounding controls. Aevum is not certified to any regulatory standard.

Who it is for

Aevum is worth evaluating if one or more of these applies:

  • Your agent accesses data about real people or organisations
  • Your agent takes actions that are difficult to reverse
  • Someone could later ask what your agent did and why
  • You operate in a regulated industry (healthcare, finance, legal, HR)
  • You have enterprise customers who ask about data governance

If none of those apply to your current project, Aevum is probably not the right tool yet. See Is It Right for You? for a detailed assessment.

How it works

Every operation passes through five unconditional barriers, requires an active consent grant, and is recorded as a signed, chained audit event before anything is written. The consent ledger tracks every grant and revocation; the episodic ledger records every operation immutably.

Read How It Works for the complete end-to-end data flow.

v0.7.4 capabilities

Zero-config developer mode. Set AEVUM_DEV=1 to get a fully working Aevum engine with no consent configuration, no database, and no signer setup. The quickstart is five minutes. The Dev to Production checklist shows what to replace before you deploy.

Eight adapters in CI. Aevum ships production-ready governance adapters for:

Adapter Package extra CI status
Anthropic Claude aevum-core[anthropic] ✓ Py 3.11–3.13
CrewAI aevum-core[crewai] ✓ Py 3.11–3.13
Google ADK aevum-core[adk] ✓ Py 3.11–3.13
LangChain aevum-core[langchain] ✓ Py 3.11–3.13
LangGraph aevum-core[langgraph] ✓ Py 3.11–3.13
MCP aevum-core[mcp] ✓ 24 round-trip tests
Microsoft Agent Framework aevum-core[maf] ✓ Py 3.11–3.13
OpenAI Agents aevum-core[openai-agents] ✓ Py 3.11–3.13

AevumOTelBridge. Routes sigchain events to any OTel backend as GenAI spans. Privacy-preserving by default: only audit_id is emitted unless content capture is explicitly opted in. Adds less than 0.5 ms p99 overhead.

Machine-verifiable conformance. An 11-invariant suite verifies the core guarantees (crisis, consent, governance, audit append-only, dual-signature, COSE receipt structure, PROV-AGENT fields, …) against any installation, reporting pass/skip/fail honestly. Run it yourself — don't take our word for it.

pip install aevum-conformance
python -c "from aevum.conformance.suite import ConformanceSuite; \
    r = ConformanceSuite().run_all(); print(r.passed_count, '/', r.total_count)"

Maintenance methodology. The Maintenance Playbook documents the four principles that govern how Aevum is developed: investigation gate, inside-out ordering, known unknowns as first-class output, and mandatory automation bias awareness at every consequential checkpoint.

Get started

pip install aevum-core
export AEVUM_DEV=1  # zero-config developer mode