Architecture 8 min
Why vector search alone isn't enough for AI memory
Single-store retrieval misses temporal context, entity relationships, and session continuity. Here's the cognitive architecture that fixes all three.
Jan 2025Read more
Now in public beta — free tier available
Give your AI agents
a real brain.
Atlas is the cognitive memory API that gives LLMs episodic, semantic, and working memory. Ingest any text. Retrieve the right context. Reason across knowledge graphs. Three API calls.
# Ingest once, remember forever
await atlas.ingest(text=my_document, user_id="acme")
context = atlas.retrieve(query="Q3 goals")
# → relevant facts, scored & ranked
await atlas.ingest(text=my_document, user_id="acme")
context = atlas.retrieve(query="Q3 goals")
# → relevant facts, scored & ranked
No credit card required · Free tier: 1,000 ops/month · SOC 2 Type II (in progress)
The ArchitectureNot just a vector store.
Not just a vector store.
A cognitive architecture.
Most "AI memory" solutions are a single Pinecone collection. Atlas implements the full cognitive memory stack — the same way the human brain stores different kinds of information in different ways.
Episodic Memory
Qdrant vector storeRaw experience chunks stored as embeddings. Your agents remember what happened — verbatim text, semantically searchable across all past interactions.
Semantic Memory
Neo4j knowledge graphStructured knowledge as a graph. Entities, relations, and multi-hop reasoning. Ask 'how is X connected to Y?' — Atlas traverses the graph to find out.
Working Memory
Redis session cachePer-session rolling context. Entity tracking, topic vector blending, hot-fact cache. Your agent knows what was said five messages ago — every time.
Memory Lifecycle
Ebbinghaus + LLM compressionMemories decay, reinforce, and compress automatically. Temporal decay, pruning below confidence thresholds, and LLM-powered cluster summarisation.
What you getEverything your agent needs.
Everything your agent needs.
Nothing it doesn't.
Six primitives. One REST API. Works with LangChain, CrewAI, LlamaIndex, or raw HTTP.
Ingest any text
POST /brain/ingest — SemanticChunker splits text, LLMGraphTransformer extracts entities and relations, stored in Qdrant + Neo4j simultaneously.
Query-adaptive retrieval
EnsembleRetriever fuses episodic + semantic results, then scores returns ranked facts with context string ready for LLM injection.
Multi-hop graph QA
GraphCypherQAChain translates natural language to Cypher, executes it against Neo4j, and returns grounded answers. No hallucination.
Memory lifecycle
Automated consolidation: Ebbinghaus decay, pruning below confidence threshold, LLM-based cluster compression into higher-order abstractions.
Usage analytics
Per-API-key monthly ops counters, retrieval precision/recall via RAGAS-style evaluation, and Prometheus metrics.
Multi-tenant safe
Every API key maps to an isolated namespace. user_id is always resolved server-side from the key — no client-supplied spoofing. Multi-tenant safe.
What builders sayTrusted by AI teams
Trusted by AI teams
shipping faster.
"We replaced a custom Redis + Pinecone setup with Atlas in one afternoon. The multi-hop graph QA alone is worth the price — our support bot now answers questions that require reading three different documents."
AM
Arjun Mehta
CTO, Synthflow AI
"The Ebbinghaus decay and automatic consolidation means our agents don't get confused by stale information. Memory management used to be our biggest headache. Now it's invisible."
PN
Priya Nair
Lead Engineer, Rephrase.ai
"Per-key namespacing is a lifesaver for B2B SaaS. Each of our enterprise customers gets fully isolated memory without any extra infrastructure. The SDK abstracts all of it perfectly."
VS
Varun Shah
Founder, AgentForge
From the teamDeep dives on
All postsDeep dives on
AI memory.
Tutorial 12 min
Building a research agent with persistent knowledge graphs
Step-by-step: use Atlas to ingest research papers, extract entities with LLMGraphTransformer, and answer multi-hop questions your LLM couldn't handle alone.
Feb 2025Read more
Engineering 6 min
The Ebbinghaus forgetting curve, implemented in Neo4j
How Atlas applies exponential decay (R = e^{-t/S}) to relationship confidence scores — and why this makes retrieval dramatically more relevant over time.
Mar 2025Read more
PricingSimple, usage-based pricing.
Simple, usage-based pricing.
Pay as your agents grow.
All plans billed monthly in INR. No hidden fees. Upgrade, downgrade, or cancel anytime.
Free
Free
1,000 ops/month
- ingest + retrieve APIs
- 1 namespace
- Episodic + semantic memory
- 384-dim embeddings (MiniLM)
- Community support
Starter
₹999/mo
50,000 ops/month
- Everything in Free
- Memory consolidation
- 5 namespaces
- Priority queue
- Email support
Need >5M ops or custom enterprise contracts? Talk to us
Talk to a founderLet's figure out if
Let's figure out if
Atlas is right for you.
We work with AI teams building production agents. Book a 30-minute call and we'll review your architecture, identify memory bottlenecks, and scope a pilot.
- Live architecture review for your use case
- Custom pricing for high-volume workloads
- Migration from existing memory solutions
- Enterprise security & compliance walkthrough
30-minute discovery call
Video call with a founder. No sales pitch. Just architecture.
Free · No commitment · Usually within 48 hours