Agent memory as data architecture

We cannot trust what the agent remembers

Memory is governed data with multiple schemas and lifecycles—not an embedding store with a prompt.

Similarity retrieval can return a relevant past fragment. It cannot tell you whether that fragment was eligible to persist, who may retrieve it, when it was true, what superseded it, or whether every derived copy was deleted.

Trustworthy memory separates recoverable session state from governed long-term records and makes scope, provenance, correction, retention, and deletion part of the schema before retrieval ranking.

See the Agent Memory as Data Architecture pillar →

Decisions that make memory trustworthy

Lifecycle: Separate exact session recovery from selective long-term reuse; they require different schemas and retention rules.
Admission: Type each durable record and capture its source, subject, authority, validity, and retention basis when it is written.
Scope: Make organization, team, user, task, and purpose boundaries structural so retrieval cannot cross them accidentally.
Change: Preserve temporal versions and supersession so the system can reconstruct what it believed when a decision was made.
Forgetting: Propagate correction and deletion through records, indexes, caches, summaries, and other derived substrates.

Read the memory path

Session state and long-term memory are different schemas — separate exact recovery from selective reuse.
Agent memory needs governance before it needs embeddings — enforce admission, provenance, scope, correction, and deletion before ranking.
Memory scope is a schema axis, not metadata — make ownership and tenancy structural.
Slowly changing dimensions already solved memory versioning — preserve what was believed when a decision was made.
Agent state needs schema migrations too — evolve stored state without silently changing meaning.

Proof and limits

Evidence status: Runnable reference build

The companion build implements a bounded local memory architecture in which governed records remain the system of record and embeddings are a derived read accelerator.

Claim: Typed, scoped, bi-temporal memory records can support provenance, correction, forgetting, and rebuildable semantic retrieval.
What exists: A public runnable implementation over DuckDB and Delta Lake with scope-filtered access, tombstone-then-purge forgetting, audit history, and self-checking scenarios.
Evidence: Inspect the agent-memory reference build
Scope and limits: This is implementation evidence for a local reference architecture. It is not deployment, compliance, benchmark, production-scale, or client-outcome evidence.

Review the memory boundary in your system

A focused architecture review traces one consequential memory path through admission, scope, temporal change, retrieval, correction, and deletion to identify the missing contracts and sequence remediation.

Review your memory architecture