Compare approaches

Compare memory architectures.

Prompts, vector retrieval, DIY systems, and Substrate preserve different units of memory—and create different trust and operating responsibilities.

Compare the approaches See Substrate's lifecycle

Four common patterns

Choose for the memory job, not the label.

Each pattern can be the right fit. Substrate focuses on durable, evidence-bounded people and organizational context across agent workflows.

Prompts and transcripts

Put recent dialogue, summaries, or selected history directly into the model context for the next run.

Short-lived continuity within a conversation or simple workflow
Fast prototyping where the context window is the main boundary

Tradeoff: Identity, freshness, conflict, and provenance are usually encoded in prose or reconstructed each time. Context size and summarization choices become part of the memory policy.

Vector retrieval

Embed source chunks or summaries and retrieve items that are semantically similar to the current query.

Finding relevant passages when exact keywords are insufficient
Adding semantic retrieval to an existing application

Tradeoff: Similarity is not, by itself, canonical identity, temporal truth, source replacement, conflict resolution, or authorization. Those behaviors need additional application logic and data models.

DIY application memory

Build a domain-specific memory layer from databases, summaries, retrieval, rules, review tools, and application permissions.

Teams with specific domain semantics and strong engineering ownership
Applications where memory is proprietary logic or needs fully controlled infrastructure

Tradeoff: The team owns the complete lifecycle: source lineage, identity resolution, status and time, correction, deletion, access scope, evaluation, operations, migrations, and user-facing trust states.

Substrate's intended architecture

Maintain replaceable source contributions, canonical entities, qualified claims, private provenance state, concise published memory, and bounded retrieval.

Agent workflows that repeatedly reconstruct people or organizational context
Managed beta use cases where evidence, change, conflict, and correction must stay inspectable

Tradeoff: This introduces a dedicated system and managed operating boundary. The current service is closed beta, connector availability varies, the correction journey is still being completed, and outcome claims remain under evaluation.

Comparison matrix

Different primitives create different responsibilities.

These are typical tendencies, not guarantees; implementations can add or omit controls.

Dimension	Prompts / transcripts	Vector retrieval	DIY memory	Substrate intent
Primary unit	Tokens, messages, and summaries in the current context	Chunks or records selected by semantic similarity	Whatever entities and records the application defines	Canonical entities and qualified published memory derived from source contributions
Identity	Usually implicit in names and prose	Often inherited from chunk metadata	Custom identity model and merge policy	Durable canonical identity with conservative resolution and explicit ambiguity
Time and change	Depends on prompt order, summaries, or application logic	Can retrieve old and new passages unless filtered or reranked	Custom temporal model	Current, historical, superseded, uncertain, and conflicting states remain explicit
Evidence	May include source text, but lineage can be lost through summaries	Commonly returns source chunks or metadata	As strong as the application's provenance model	Published memory retains bounded citations while detailed evidence stays private
Correction	Edit the prompt, summary, or upstream transcript selection	Update or delete indexed chunks and dependent summaries	Custom review, replacement, and history flow	Replace source generations and recompute projections; complete customer self-service remains beta work
Authorization	Set by what the application places in context	Requires filters and access checks around retrieval	Custom roles, scopes, and enforcement	Retrieval is intended to be bounded by named subject, agent, scope, and task; memory grants no action authority
Operational burden	Low infrastructure, rising prompt and summarization complexity	Index, embedding, filtering, freshness, and evaluation operations	Full engineering and product ownership	Managed dedicated memory system with an evolving beta interface and deployment boundary

Vector search and language-model context can still be part of a canonical memory architecture. The distinction is whether they are the complete memory model or components inside a wider identity, evidence, time, scope, and correction system.

Substrate's intended architecture

One published memory surface, with private state for the work behind it.

Replaceable source evidence and private workflow state support canonical, qualified pages retrieved for a named scope and task.

Replaceable contributions

A successful reprocessing generation replaces one source's prior active contribution without deleting unrelated support.

Canonical identity and relationships

Repeated mentions can resolve to durable entities while ambiguous people and organizations remain separate or unresolved.

Qualified publication

Safe supported context is projected into concise pages with time, state, conflict, and provenance kept explicit.

Bounded retrieval

An agent receives permitted published context for the current subject and task—not raw events or the private ledger.

Decision guide

Start with what must stay true after the conversation ends.

Compare trust and operating requirements, not feature labels.

Do you only need continuity inside a short interaction?

Prompt or transcript context may be the simplest sufficient choice. Add summarization carefully and decide what happens when the context window fills.

Do you mainly need to find relevant source passages?

Vector or hybrid retrieval may be a strong fit. Add metadata filters, authorization checks, freshness handling, and citation behavior required by the application.

Is memory core proprietary application logic?

A DIY system can provide full control. Budget for the full lifecycle—not just storing summaries—including identity, correction, deletion, evaluation, and operations.

Do people, evidence, change, and scope need a durable home?

Substrate may be worth evaluating when agents repeatedly reconstruct the same relationships and the memory itself must be inspectable. The current fit is a managed closed beta, not a self-serve or generally available service.

Evaluate with a real workflow

Bring the memory problem, not a predetermined architecture.

We will review the source boundary, identity and change requirements, inspection needs, application authority, and evidence for continuing the beta.

Apply for closed beta Developer model