Seven US companies lead agent memory; five ship graph memory, but only Zep and Cognee document bi-temporal validity
Asked (summary):
Treat agent memory / agentic retrieval as the vertical. Find the top 7 US companies and extract 7 comparable details each — company, product, US HQ city, memory approach, storage model, temporal memory, relational/graph memory, and public pricing or latest disclosed funding — with a primary source URL on every cell, cross-compared in a correctable table.
Seven companies, one row each, ranked by live-web prominence — distinct source mentions in a broad vertical scan that merged 8 diverse searches into 1,357 unique pages and extracted 612 company/product rows. This is not a revenue, valuation, or customer-count ranking. Six of seven publish pricing; Engram is represented by its $98M seed. Unresolved cells stay blank and are flagged, not invented.
Comparison matrix — one row per company, ranked by web prominence
native / strong documented as core capability
limited / filtering partial capability per docs
undocumented not publicly resolved
Source links: blue = official docs/pricing, amber = secondary reporting
| Company & product | Location | Architecture | Capabilities | Commercial signal |
|---|
| Company / product | Prominence
(web mentions) | US HQ city | Memory approach & storage | Temporal memory | Graph / relational memory | Pricing or latest funding |
Prominence signal at a glance
Cross-company capability summary
Zep
Konig purpose-built graph database; Context Graph of facts, entities, episodes
Bi-temporal: four timestamps per edge, point-in-time queries
Cognee
Pluggable relational + vector + graph stack (SQLite/Postgres, LanceDB, Ladybug/Kuzu/Neo4j)
Bi-temporal validity stamps; superseded facts kept, not deleted
Mem0
Hybrid vector + entity graph + SQL audit log
Temporal reasoning layer, on by default (92.0 LoCoMo temporal)
Vectorize (Hindsight)
PostgreSQL memory graph with consolidation and typed causal edges
Staleness/currency tracking, not bi-temporal
Letta
Git-backed MemFS: markdown memory files in a real git repository
Datetime-filterable passages only
Supermemory
Container/room graph; access follows graph position
Date/timestamp query filters
Engram
Not publicly resolved — do not overstate on funding coverage alone
Not publicly resolved
Findings
Five of seven document graph/relational memory — Mem0, Supermemory, Hindsight, Zep, Cognee — while Letta's graph support is undocumented in its own materials.
docs.mem0.aiTemporal depth splits sharply: Zep's edges carry four timestamps for bi-temporal point-in-time queries, while Supermemory and Letta offer only timestamp filtering.
blog.getzep.comHindsight is the only leader keeping its entire memory graph in plain PostgreSQL tables — entities, junctions, and co-occurrence edges — with no external graph database.
hindsight.vectorize.ioEngram is the only company with no public pricing or architecture disclosure; its $98M seed at a $600M valuation (June 2026) is its sole public commercial signal.
finsmes.com
Method: shortlist of 7 US agent-memory companies ranked by live-web prominence — distinct source mentions in a scan that merged 8 searches into 1,357 unique pages and yielded 612 company/product rows; not a revenue or valuation ranking. One row per company with per-field source URLs; long architecture text truncated in the overview and expandable in place. Blank cells mean the fact was not publicly resolved; amber-marked links are secondary reporting where no primary statement was found (Zep HQ, Engram funding/architecture, Vectorize HQ, Letta graph note, Cognee pricing note). Extracted from public docs, pricing pages, and press coverage.