The associative memory database
Store patterns. Learn what they share. Recall from partial cues. HeatherDB runs on your CPUs, through HTTP or directly in Rust.
Related observations shape shared memory. Connections shown illustrate associations in encoded data.
Install the engine, store two patterns and recall from an incomplete cue.
Follow the Quickstart ↗curl -fsSL https://heatherdb.com/install.sh | shThe scripts select your platform and verify its checksum. Inspect scripts and all builds ↗
Explore working examples of associative memory and vector algebra. Each lab shows the API calls behind its result.
Coffee · $4.50 Groceries · $62 Lunch · $14
Overseas purchase · $4,200
A different pattern stands out.Explore how encoded spending patterns can surface an unusual event. A demonstration, not a validated fraud detector.
Open the live labEach lab runs against HeatherDB. Open its inspector to see the memory and the API calls behind the result.
Inside your application
Your application supplies vectors. HeatherDB stores and recalls patterns. Your application decides how to use the result.
Turn records, events or content into vectors with an encoder that fits your task.
Adapt shared memory or preserve separate locations. Read through HTTP or Rust.
Use reconstructed patterns or query documents with the APIs designed for each.
Associative reconstruction returns a vector. Document lookup and identity use separate APIs.
Explore the architecture ↗Two ways to write. Associative recall in both.
Related observations can update a shared representation. Novel patterns receive their own locations.
Append a separate vector location for every observation, without adapting earlier locations.
Choose the mode when creating a database. Compare the behaviours and trade-offs ↗
CPU measurements with source revisions, workloads and raw samples. Performance and memory quality are evaluated separately.
128 dimensions · 200 prefill writes.
Criterion mean on Apple M4.
| Workload | Competitive | Plain |
|---|---|---|
| One vector · one call | 4.72 ms | 4.39 ms |
| 100 vectors · individual calls | 460.64 ms | 368.01 ms |
| 100 vectors · one batch | 9.92 ms | 4.19 ms |
20 paired samples, synchronous durability. Single-write timing does not establish a clear difference.
Intervals, limits and reproduction ↗Know how it runs, what it guarantees and where the boundaries are.
The engine runs on CPUs. Choose a native binary, a container or embedded Rust.
Read the details ↗Synchronous LMDB commits by default, with documented backup and restore procedures.
Read the details ↗Database-scoped users and authenticated HTTP operations. Put exposed HTTP behind TLS.
Read the details ↗One engine process per data directory on local block storage. Replication is outside v1.
Read the details ↗Learning and durable commits have a cost. Batch writes and choose the memory mode your task needs.
Read the details ↗AGPL-3.0-only. Source, release artifacts and contribution guidelines are public.
Read the details ↗Start locally. Inspect the source. Make something useful.