PILLAR · 01
Records & Provenance
Every record is atomic, typed, and cryptographically sealed at ingestion.
We call it the Universal Provenance Layer (UPL). Each record
carries a source pointer, a parser version, the user's consent scope, and a
signature. Re-run the source ten years from now and the signature still verifies.
Nothing is a blob. Nothing is a summary. Every fact is provable.
Tech: Ed25519 signatures · canonical JSON · content-addressed storage · per-record TTL + revocation · audit log on every read.
Deep dive →
PILLAR · 02
Memory Stack
Episodic memory ("on March 12 you bought…") flows into semantic memory ("you usually
buy gas every 10 days") flows into a queryable graph. The stack is a real stack —
structured records at the bottom, embeddings in the middle, an
entity graph in the same Postgres on top. Every answer at the top is
grounded to a record at the bottom.
Tech: Postgres + pgvector + entity graph on a single tenant-isolated cluster. Bi-temporal queries. Recall is graph-first, embedding-second.
Deep dive →
PILLAR · 03
Schema-aware Connectors
Connectors aren't dumb pipes. Each one is a schema — Gmail receipts parse into
charges, Plaid statements parse into line items, Tesla telematics parse into trips,
iMessage parses into typed conversations with entities resolved. The hard work is
the parser, not the OAuth flow. We've written dozens and add roughly one a week.
Tech: Schema-versioned parsers · idempotent backfill · incremental sync · per-connector consent scope · self-healing on schema drift.
Deep dive →
PILLAR · 04
Time-aware Chronology
Most personal databases think in records-as-strings. We think in records-as-events
with both occurred-at and known-at timestamps —
so a charge that posts three days after the swipe doesn't break the chronology, and
an amended tax return doesn't overwrite the original. Bi-temporal from day one. Audits
and discovery become trivial.
Tech: Bi-temporal model · event-sourced corrections · "as of date X" replay · timeline reconstruction in O(log n).
Deep dive →
PILLAR · 05
Agent API
We don't want to be the chatbot — we want to be the layer every chatbot trusts.
Lossless speaks MCP and REST, with per-scope consent
(read receipts but not messages; read vehicles but not finance) and per-tool rate limits.
Claude, GPT, Gemini, and Apple Intelligence can all read from the same ledger — the user
decides who and what.
Tech: MCP server · OAuth 2.1 + PKCE · scoped tokens · per-tool consent matrix · every read written to the audit log.
Deep dive →
PILLAR · 06
Voice → Typed Records
Speak it. We type it. We seal it. Voice notes flow through the same parser pipeline
as connectors — an entity-resolved typed record at the bottom, indexed and graph-linked
at the top. The transcript itself is never the source of truth — the typed record is.
Tech: Whisper-v3 + speaker diarization · entity resolution against the user's existing graph · PII redaction before storage.
Deep dive →