A database that adapts to your data. Each write reshapes its memory, building associations as new information arrives. Run intelligence on everyday CPUs, without a separate model-training pipeline.
Stored data can recommend, spot unusual activity and reason by analogy. Explore six live demos, without a separate model or GPU.
Coffee · $4.50 Groceries · $62 Lunch · $14
Overseas purchase · $4,200
A different pattern stands out.Learn the pattern in everyday spending, then flag a charge that falls outside it. No hand-written fraud rules.
Open the live labEach lab runs against HeatherDB. Open its inspector to see the memory and the API calls behind the result.
HeatherDB learns associations as you write data. Use that memory for recommendations, pattern completion and anomaly detection, without a separate training pipeline.
As new data arrives, the memory adapts in place. Connect your app through HTTP, using your own vectors or embeddings.
Build your first memory Explore the HTTP APIPrefer a binary? Download for your platform, or use HeatherDB directly as a Rust library.
docker run -d --name heatherdb \
-p 127.0.0.1:6380:6380 \
-v heatherdb_data:/var/lib/heatherdb \
-e HEATHER_DIMENSION=8 \
-e HEATHER_ADMIN_USER=admin \
-e HEATHER_ADMIN_PASSWORD='quickstart-password' \
aphorion/heatherdb:latestcurl -u admin:quickstart-password \
-H 'Content-Type: application/json' \
-d '{"vectors":[[1,1,1,1,0,0,0,0]]}' \
http://localhost:6380/collections/quickstart/writecurl -u admin:quickstart-password \
-H 'Content-Type: application/json' \
-d '{"query":[1,0,0,-1,0,0,0,0]}' \
http://localhost:6380/collections/quickstart/readThe read returns a reconstructed vector, not a stored row. This example binds to localhost and keeps data in a named volume. Use separate credentials and a pinned release for deployment.
HeatherDB is read-optimized. Writes organize and persist the memory so reads can stay fast. Measured on an Apple M4 with 16 GB RAM.
| Operation | 64 dimensions | 128 dimensions | 384 dimensions |
|---|---|---|---|
| Single write | 5.09 ms | 5.23 ms | 5.27 ms |
| Iterative read | 17.47 µs | 32.39 µs | 120.07 µs |
Compared with 526.41 ms for individual writes at 128 dimensions: approximately 42.3× faster in this run.
Full results & methodology Compare learning on and offMeasured 2026-11-05, HeatherDB 1.0.0. Means from 20 Criterion samples, using current engine defaults (data-seeded locations, k = 20). Native engine calls exclude HTTP, authentication, network and encoding. Speed does not measure associative recall quality.
cargo bench --locked -p heather_db --bench read_latency -- read/dimension --sample-size 20 --warm-up-time 1 --measurement-time 2Persistent storage, access controls and recovery tools around the associative memory. Open source under AGPL-3.0, running on hardware you control.
LMDB-backed persistence on local disk. Run the HTTP server or embed the same engine in a Rust application.
Server or embeddedMultiple databases and collections, with vector dimensions and memory settings configured per database.
Manage databasesAuthentication is on by default. Give each application a database-scoped user instead of sharing admin credentials.
Users and permissionsAttach JSON metadata to vectors. Fetch documents by ID, query candidates by similarity, and delete records through the API.
Document APITake live snapshots per database. Use cold backups and restores when the engine is stopped.
Backup and restoreHealth and readiness probes, collection statistics, and per-database audit logs give operators a view into the engine.
Deployment checklistHeatherDB 1.0 establishes the documented API and storage compatibility baseline. Vector dimension is fixed per database, and each data directory must have one engine process on local storage.
Associative reads can blend memories as a collection fills. Keep exact records through document APIs and evaluate recall on your own data. Deleting a document removes it from retrieval; removing its learned influence requires rebuilding the collection.
Terminate TLS at your proxy. There is no built-in Prometheus endpoint yet; use collection statistics, logs and audit records.
Read the boundary conditionsHeatherDB is open source and runs on a laptop. Install it or explore the live demos. No signup needed. Licensed under AGPL.
Contributions are welcome, including AI-assisted work. Every pull request is reviewed manually: explain your changes, document their behaviour and provide checks you have run. Read the contributor guide.