Semantic graph memory that remembers correctly
A privacy-first, higher-accuracy memory layer for AI agents. Hybrid retrieval, automatic dedup, and an extract-and-consolidate pipeline — self-hosted or managed, without the vendor lock-in.
from memtavo import MemtavoClient
client = MemtavoClient(api_key="YOUR_API_KEY")
client.add(user_id="user_123", messages=[
{"role": "user", "content": "I'm allergic to peanuts"}
])
results = client.search(
user_id="user_123",
query="What should I avoid eating?"
)
Built to beat the accuracy ceiling
Benchmarked head-to-head on LoCoMo and LongMemEval against the leading memory layers.
| Benchmark | memtavo | Mem0 | Zep |
|---|---|---|---|
| LoCoMo | 96.4% | 91.6% | 80.3% |
| LongMemEval-M | 96.0% | 93.4% | — |
Retrieval built for accuracy, not just recall
Every layer of memtavo's pipeline exists because a simpler version of it missed real queries in testing. Here's what's actually running underneath.
Hybrid retrieval
Vector similarity
Asymmetric embedding model tuned for query-to-document matching, not generic sentence similarity.
Lemmatized keyword overlap
"ran" matches "run", "children" matches "child" — catches exact-term queries that pure vector search misses.
Graph-augmented search
Topic-graph traversal surfaces related memories that share context but not vocabulary.
Multi-rule deduplication
Source-aware skip
Conversational turns bypass dedup — they're inherently unique.
Exact-match fast path
Normalized text lookups catch verbatim duplicates in O(log n).
Hybrid gate
Requires high cosine similarity and keyword overlap before merging — fewer false positives.
Length-aware thresholds
Short entries use stricter thresholds than long ones, where coincidental similarity is more likely.
Extract + consolidate
Extraction
Batched LLM extraction turns raw turns into atomic, timestamped facts — run on demand or scheduled.
Consolidation
Similar facts are clustered and merged into clean statements, keeping memory tidy as it grows.
Optional knowledge graph
Triple extraction and contradiction detection for teams that need multi-hop reasoning.
How it works
Store
Send conversation turns or facts through the API — self-hosted or managed.
Extract & consolidate
Atomic facts get extracted, deduped, and merged into clean long-term memory.
Retrieve
Hybrid retrieval returns the most relevant memories, ranked and ready for your prompt.
Pricing that scales with you
Start free. Upgrade when you outgrow the tier — not before.
Free
Try memtavo on a personal project
- 1,000 retrievals/mo
- 500 memories
- REST API access
- 1 user
- Community support
Starter
For small teams shipping their first agent
- 10,000 retrievals/mo
- 10,000 memories
- 5 users
- Email support
- Basic analytics
Pro
For production workloads at scale
- 100,000 retrievals/mo
- 100,000 memories
- Unlimited users
- Priority support
- Python & JS SDKs
- Webhooks
- P95 < 300ms SLA
Enterprise
For organizations with compliance needs
- Unlimited retrievals
- Unlimited memories
- Dedicated infrastructure
- SSO / SAML
- SOC 2, audit logs
- On-prem option
- P95 < 200ms SLA
Need more than what's listed here? Talk to us about a custom plan.
We're not trying to out-integrate Mem0
We're building for the developers who tried an existing memory layer and hit its limits — dedup that lets duplicates through, accuracy that drops on long conversations, and pricing that jumps straight from a free tier to hundreds a month.
Accuracy first
Every retrieval decision — hybrid signal fusion, dedup thresholds, consolidation rules — was tuned against real benchmarks (LoCoMo, LongMemEval), not shipped on intuition.
Privacy and control
Self-hosted gets the full pipeline with no usage limits. Cloud is the same API with zero setup, multi-tenancy, and managed scaling — you choose where your data lives.
Honest pricing
A real tier between free and enterprise, so a small team doesn't have to choose between a 500-memory sandbox and a five-figure contract.
Where we stand today
We're early. Our integration ecosystem is smaller than the incumbents' — we're focused on getting a few frameworks (LangChain, LlamaIndex) right rather than shipping twenty shallow ones. What we won't compromise on is retrieval accuracy and dedup quality, because that's the part that actually breaks agents in production.