Structured memory
Keep facts, records and decisions in managed tables that agents query with SQL.
LeanMemory
Memory that learns, not just a vector store.
A vector database returns what matches. LeanMemory returns what it has learned to be right: it studies the questions your agents ask, keeps the answers that proved correct, and gets more accurate the more it is used. Structured records, unstructured knowledge and the full history of events live side by side, under one key.
01
One memory, every shape of data
Tables for facts, vectors for meaning, events for history and files for everything else, unified under one workspace and one key.
02
Learns from outcomes
Tell LeanMemory which results helped. Feedback shapes what it surfaces next, so recall improves the more your agents work.
03
Private by design
Every workspace gets its own memory, created on first use. Nothing crosses accounts, and nothing is deleted when a balance runs low.
Public benchmark · September 2026
How it was measured
Level with state-of-the-art RAG on first sight. Ahead once it has learned.
Five public question sets from the HippoRAG 2 evaluation, 1,000 questions each. Every system used the same embedding model, reranker, reader and judge, so the only difference was how it found the passages, and whether it learned from use.
First sight
62.4
Conventional RAG 58.5, state-of-the-art RAG 62.2. Same models, same questions, nothing learned yet.
After studying the questions
72.1
Two study rounds over the questions it was asked, and nothing else: no answers, labels or gold passages.
After being taught
87.5
Reference answers for the questions it still missed, given the way a person would teach it. 85.7 on rewordings it had never seen.
Score by question set
The average of answer accuracy and recall at 5, in per cent. Each LeanMemory bar starts at its first answer to every question and adds what it gained from studying the questions and from being taught. Conventional and state-of-the-art RAG do not learn, so their bars are their first answers.
Conventional RAG
State-of-the-art RAG
LeanMemory
first sight
after studying the questions
after being taught
Show the figures as a table
| Question set | Conventional RAG | SOTA RAG | First sight | Study 1 | Study 2 | Taught |
|---|---|---|---|---|---|---|
| NQ simple | 63.2 | 65.2 | 64.8 | 69.2 | 69.2 | 84.3 |
| PopQA simple | 58.2 | 58.0 | 57.9 | 61.6 | 61.7 | 76.8 |
| MuSiQue multi-hop | 35.3 | 42.9 | 43.5 | 54.5 | 55.3 | 83.6 |
| 2Wiki multi-hop | 59.0 | 59.6 | 59.9 | 81.9 | 82.7 | 94.4 |
| HotpotQA multi-hop | 76.9 | 85.1 | 86.0 | 91.6 | 91.6 | 98.4 |
| Average | 58.5 | 62.2 | 62.4 | 71.7 | 72.1 | 87.5 |
In per cent. SOTA RAG is the state-of-the-art recipe. The last four columns are LeanMemory: its first answer, after one and two study rounds, and after teaching. Neither RAG system learns, so each has one figure, its first answer.
Finding the passages takes no model call: LeanMemory answers in 453 ms at the median, the state-of-the-art recipe in 1.6 s. Plain vector search is faster still at 26 ms, and scores lowest.
See the full benchmark
Capabilities
What your agents can do with LeanMemory.
Semantic recall
Store embeddings from the model you prefer and retrieve by meaning, with metadata filters.
Event memory
Append-only, time-ordered streams capture everything that happened, ready to query and replay.
File memory
Private storage for documents, media and artifacts, shared through signed links.
Bulk in, bulk out
Load and export millions of rows over a high-throughput columnar channel, or connect with a native driver.
Learns from feedback
Signal which results helped and LeanMemory adapts what it brings back next time.
How your agent uses it
SQL, vectors and events, one connection.
Agents read and write LeanMemory with SQL, semantic search and event queries on the same endpoint and key as the rest of LeanTools, or through a native driver for bulk work.
API reference-- What happened with this customer lately? SELECT occurred_at, channel, summary FROM customer_events WHERE customer_id = 'C-1042' ORDER BY occurred_at DESC LIMIT 5;
Use cases
Where teams put LeanMemory to work.
Context that follows every conversation
Support and sales agents pick up exactly where the last interaction ended.
Decisions with a paper trail
Agents record what they decided and why, so reviews and hand-overs take minutes.
Operations you can replay
Event history turns every run into something you can audit, analyse and learn from.
What kinds of data can LeanMemory hold?
Structured records in tables, embeddings for semantic recall, time-ordered event streams and files of any type, all in one workspace and reachable with one API key.
Do I bring my own embeddings?
Yes. LeanMemory stores and searches the vectors you produce with the embedding model of your choice, so recall stays consistent with the rest of your stack.
Is our memory isolated from other customers?
Every workspace has its own memory. A request that names data outside your workspace is refused before it ever reaches storage.
What happens if our balance runs out?
Nothing is deleted. Your memory stays intact and new writes pause until you top up.
How is LeanMemory priced?
By data held per GB-month, plus compute time and data moved for queries and loads. Management calls are free.
How does LeanMemory compare with RAG?
On the five public question sets of the HippoRAG 2 evaluation, 1,000 questions each, with the same embedding model, reranker, reader and judge for every system, LeanMemory scored 62.4 on first sight against 62.2 for state-of-the-art RAG and 58.5 for conventional RAG. After studying the questions it was asked it reached 72.1, and 87.5 once taught the answers it still missed. The full results and the method are on the benchmark page.
Better together
Pairs well with LeanMemory.
LeanData
Datasets your agents can discover, query and subscribe to, through one consistent interface.
LeanModel
OpenAI-compatible text, image, speech, music, video, avatar and realtime voice models, metered on the same credits as the rest of your stack.
LeanConnect
Connect agents to the apps your business runs on, with credentials kept safe and every write under your control.
Start building with LeanMemory.
One endpoint, one key and every LeanTools product, from your first call to production.