The Memory Governance Manifesto

AI memory — the persistent data stored by RAG systems and agents — has become a core security and compliance surface. Yet it remains largely unguarded, with no standard practices for what can be stored, retrieved, or forgotten.

AI memory is a lifecycle

Write

Store

Retrieve

Unlearn

Audit

Memory doesn't just happen — it flows through distinct stages:

  • Write:Data enters the system (user input, document upload, API call)
  • Store:Embeddings, chunks, or raw data persist in vector DBs, SQL, or logs
  • Retrieve:Queries return context for LLM prompts or agent actions
  • Unlearn:Data is deleted on request (GDPR, customer policy, security incident)
  • Audit:Every action is logged with proof of compliance

Each stage requires active governance, not passive logging.

Why governance differs from observability

Observability tells you what happened. Governance controls what can happen.

Logs show you that a user's SSN was stored. A governance layer would have blocked it before write, or redacted it automatically. Observability is diagnostic. Governance is preventative.

You can't audit your way to compliance if the violations already happened. You need enforcement at the data plane.

The unguarded attack surface

Without memory governance, AI systems expose critical vulnerabilities:

PII leakage

User uploads contain SSNs, credit cards, health records — all vectorized and retrievable

Secret persistence

API keys, passwords, and tokens embedded in documents or chat logs

Unauthorized retrieval

Agents query across user boundaries, returning data they shouldn't access

Failure to forget

GDPR deletion requests leave orphaned vectors, chunks, and metadata

These aren't edge cases. They're inevitable in production RAG and agent systems.

What 'provable forgetting' means

Deleting data is easy. Proving it was deleted — everywhere — is hard.

Provable forgetting requires:

  • A manifest of what was stored (chunks, embeddings, metadata)
  • Deletion jobs that span all storage backends
  • Counts: N records deleted, M vectors removed, X chunks cleared
  • Audit records with timestamps, affected IDs, and job status

Without this, "we deleted it" is just a claim. With it, it's a certificate.

Reference architectures

RAG SaaS

Multi-tenant RAG product (document chat, search, Q&A)

Flow: User upload → Nightwatch (detect PII, apply policy) → Vector DB

Retrieval: Query → Nightwatch (filter by tenant, redact) → LLM

Unlearning: Delete user → Nightwatch (remove all chunks/vectors) → Audit log

Regulated enterprise

Internal AI assistant for finance/legal/healthcare

Flow: Employee doc → Nightwatch (classify, fingerprint) → Postgres + Vector DB

Policy: Block PHI outside compliance team; redact card numbers always

Audit: SOC 2 / HIPAA logs with retention rules and access proofs

Where Nightwatch sits

Nightwatch is a governance layer, not a replacement for your stack. It sits between your applications and memory stores:

  • Apps and agents call Nightwatch APIs instead of direct DB writes
  • Nightwatch runs detection, applies policy, stores metadata, and forwards data
  • Retrieval goes through Nightwatch for scoped, filtered results
  • Unlearning jobs coordinate across all backends and produce receipts

You keep your vector DB (Pinecone, Weaviate, Qdrant). You keep your SQL database. Nightwatch just governs what flows through.

Join the design partner program

Help define memory governance for AI systems. Get early access and shape the category.