Session-level. Fast. Temporary.
The current conversation context. Active entities being discussed. Cross-operation context shared between agents inside one session. Token-budget aware.
TTL: SESSION DURATION · AUTO-CLEANUP
Home / Platform / Memory Stack
Platform · How records become memoryA real stack. Not a context window.
Most "AI memory" is a sliding chat-history window. Lossless's memory is a literal three-layer stack: structured records at the bottom, embeddings in the middle, an entity graph on top. All three layers sit in the same RLS-protected Postgres as your records. Recall is graph-first, embedding-second — and every answer cites the exact record the answer came from.
← One tenant, one Postgres
Records → embeddings → graph
Recall is graph-first →
What memory means here
Not chat history. Real, indexed records of your life.
When most AI products say "memory," they mean a sliding window of your conversation. Lossless means something engineering-grade: every email, statement, drive, document, and voice note is a structured record — typed, dated, linked to the right entities, and retrievable in graph order. Every answer cites the exact record it came from.
- ✓ Records, not just embeddings — typed and queryable
- ✓ Every record cites the source, with a link back to the original
- ✓ Backfilled from years of email + accounts on day one
- ✓ Zero loss on update — bi-temporal, old versions are kept and dated
- ✓ Memory survives model upgrades, app rewrites, and re-architectures
Memory also has a temporal life-cycle
Working → Episodic → Semantic.
The three-layer stack above is the storage architecture. This is the temporal one — borrowed from cognitive science. Memories evolve from fast and disposable, to recent and structured, to permanent and consolidated.
Recent interactions. 30–90 days.
Every interaction stored with vector embeddings. Entities extracted and linked to the graph. Each episode gets an importance score (1–10). High-importance episodes auto-promote to Semantic.
PROMOTION: HIGH-IMPORTANCE → SEMANTIC LAYER
Long-term facts. Permanent.
Consolidated knowledge facts. User preferences and patterns. Learned procedures. Cross-domain generalizations. Each fact carries a confidence score (0.0–1.0) and full source-tracking provenance.
CONFIDENCE: 0.0–1.0 · SOURCE-TRACKED
The Memory Flow Pipeline runs continuously: working memories crystallize into episodes; episodes consolidate into semantic facts; facts decay or strengthen with confidence over time. The user owns the entire pipeline.
How it's built
Four ideas that make this work where chat-history doesn't.
Typed records
Every memory item has a schema — vehicle service, email, statement line, place visit. Schemas mean retrieval can be precise instead of fuzzy.
Linked, not flat
Records reference each other: a service event links to its confirmation email, its invoice PDF, the vehicle, the service center, the bank charge.
Dated & versioned
Every record has when-it-happened and when-it-was-recorded. Updates create new versions; nothing is silently overwritten.
Local-first export
Your full memory is exportable as JSON + original files at any time. The store is the durable artifact — Lossless is the application that reads it.
"Other 'memory' products were a nice gimmick — they remembered what I told them. Lossless remembers what I did. That's a completely different product."
Continue the architecture tour
You've seen the layers. Now see the top plate up close.
Next pillar: the entity graph itself — every record a node, every relationship a typed edge, every traversal sealed back to a source record.