The brain / memory architecture

A real memory layer.
Not a chat log.

Brains turns your inbox, drive, calendar, calls, and conversations into a typed, queryable substrate, indexed by vector embeddings, full-text search, and a knowledge graph.

Hybrid retrievalKnowledge graphYour data + keys20-50ms retrieval
Brains memory architecture diagram with typed surfaces around the brain logo
Ingest

One pipeline. Every source.

Connect Gmail, Drive, Calendar, Slack, GitHub, or your own source. Brains backfills history, then keeps it fresh through file watchers, polling, native pulls, and authenticated webhooks.

One event shape

Every source emits a kind, URI, content hash, and ingestion timestamp before downstream work begins.

No duplicates

A 24-hour content-hash window prevents retries and overlapping sources from ingesting the same event twice.

Reliable by default

The supervisor validates, deduplicates, writes provenance, and dispatches chunking, embedding, and entity jobs.

  • Real-time watcherAny local folder you point at Brains.
  • Polled inboxEmail and shared drives, with retries on failure.
  • Webhook receiverAuthenticated push from connected integrations.
  • Native pullsGmail, Drive, Calendar, Slack, GitHub, and more on a per-source cadence.
Store

Typed pages. Not a blob.

Every item becomes a structured page with frontmatter, a normalized body, an extracted timeline, and full provenance. Pages are the unit of retrieval, sharing, and reasoning, and the schema stays open to your own page kinds.

Image placeholder - typed page anatomy
  • Compiled truthA normalized body that receives a 2x retrieval boost.
  • TimelineAn extracted temporal spine from event dates, metadata, and filenames.
  • Open frontmatterStructured fields that can inform ranking and filtering.
  • Many indexesVector, full-text, and multimodal passages generated from the same page.
Connect

A knowledge graph, extracted for free.

People, companies, deals, and events become typed relationships such as attended, works at, invested in, founded, and advises. The graph powers expert discovery, trajectory questions, and entity-aware retrieval.

Entity graph linking people, companies, spending, and related events
  • Typed edgesThe relationship kind is part of the schema, not guessed at query time.
  • Walked at retrievalFind experts, trace trajectories, and expand related code or work.
  • AuditableEvery edge points back to a source page, so nothing arrives without provenance.
Retrieve

Three signals. One ranked answer.

Vector embeddings, full-text search, and graph walks run in parallel. Reciprocal Rank Fusion merges them, then compiled truth, backlinks, salience, recency, and an optional cross-encoder sharpen the result.

Vector

Semantic similarity across chunked text and multimodal content.

Full text

Weighted, stemmed, and exact-match aware keyword retrieval.

Graph

Typed entity relationships walked to depth two, with auditable sources.

20-50msTypical brain retrieval
K=60RRF rank constant
2xCompiled-truth boost
3 lanesRun in parallel
Protect

Your data. Your keys. Your brain.

Brains is architected so your memory stays inside the trust boundary you set. Row-level security, scoped tokens, private fences, reversible deletion, and generation-aware caches make privacy and freshness part of the system itself.

Row-level security

Scoped at the storage layer.

Every read carries a source ID resolved from the caller token. Cross-source enumeration is blocked at the API layer, not just discouraged.

Scoped tokens

OAuth clients see what you allow.

Tokens carry allowed sources, so a client scoped to shared pages cannot read private memory, even if it asks the right question.

Private fences

Remote callers get a stripped view.

Remote MCP requests skip private facts and per-token allow lists entirely. Local tools keep the access you explicitly grant.

Never trains

Your pages do not train anyone's model.

Brains never sends user pages to train external models. Embedding and synthesis calls run on your keys, with prompts you can audit.

Soft delete

Deletes are reversible for 72 hours.

Pages marked for deletion stay recoverable during the grace window before autopilot permanently removes them.

Generation clock

Caches cannot go stale.

A global and per-page generation clock invalidates semantic caches the instant a page changes, so reads stay fresh.

The memory layer your AI deserves.

Join waitlist →