Skip to main content

Atlas - Cognitive Memory

Atlas gives AI agents persistent, queryable memory — the same way the human brain stores different kinds of information in different ways. Most “AI memory” solutions are a single vector collection. Atlas implements the full cognitive memory stack across three layers:
LayerStoreWhat it remembers
EpisodicQdrant (vector)Raw experiences — what happened, verbatim
SemanticNeo4j (graph)Structured facts — entities, relations, ownership
WorkingRedis (cache)Session context — what was said 5 messages ago

Why not just stuff context into the prompt?

Prompt-stuffing doesn’t scale:
  • Token limits force you to truncate, losing older facts
  • All facts are weighted equally — a stale fact competes with a fresh one
  • No concept of which facts are still true
  • Cost grows linearly with memory size
Atlas retrieves only the most relevant, most recent context for each query — typically 5–15 facts instead of thousands of tokens.

The scoring function

Every retrieval is ranked by a hybrid score using following parameters:
SymbolSignalDefault weight
VSemantic similarity to your query0.20
RRecency (exponential decay)0.70
FAccess frequency0.10
AGraph relation alignment0.50
Recency dominates by design. A fact stored today outranks a contradicting fact from last month — automatically, with no configuration.

What you can build

  • Support agents that remember every ticket, preference, and resolution
  • Research assistants that accumulate knowledge across sessions
  • B2B SaaS agents with fully isolated memory per customer namespace
  • Coding agents that remember architecture decisions and tech choices
  • Personal AI that learns your preferences over time

Base URL

https://api.atlas.bsyncs.com All endpoints require an X-API-Key header. Get yours at atlas.bsyncs.com/signup.