01 / Local data and agent memory
Hyphae
A local data engine and shared memory for coding agents, with retrieval and verifiable records.
Stable 3.0.0
Explore this softwareRust / Apache-2.0
Open systems research / Est. 2026
Systems that remember.
Research you can inspect.
We build open infrastructure for durable data, agent memory, and research with a verifiable record. Explore what works today, what we have learned, and where we can build together.
Reviewed on
Hyphae Research Foundation
Hyphae / Agent Memory
Useful context, beyond one session.
Agent Memory gives coding agents a shared local record of decisions, constraints, and reusable knowledge. Recall stays scoped to the project and can carry evidence of the stored snapshot.
Discover Agent MemoryAgent Memory
- 01Save useful context
- 02Recall by project
- 03Verify the record
Hyphae Local data and agent memory
Open software
Four projects. A shared commitment to evidence.
Four open projects connect durable data, observable training, controlled experiments, and CPU inference.
02 / Training instrumentation research
Hytorch
A research prototype that records internal training writes, making them auditable and replayable.
Research prototype / preprint 1.1
Explore this softwareRust + Python / Apache-2.0
03 / Controlled transformer research
Hyphae Transformer
Controlled transformer experiments and bounded controllers over frozen models.
Public alpha 0.2.0
Explore this softwarePython + PyTorch / MIT
04 / Exact ternary CPU inference
Hyphae BitNet
CPU inference for supported ternary models, with exact integer contracts and explicit memory budgets.
Runtime 0.3.2 / release tag v0.3.3
Explore this softwareC/C++ + Python / Apache-2.0
Ecosystem integration
Hyphae Memory for Omarchy
Hyphae Memory brings project memory to the Omarchy desktop through an independently managed local service.
Client 0.2.0 / separate service requiredRecent work
Public progress, with its evidence.
Hyphae 3.0.0
Expanded SQL, keyspace, and search capabilities; published packages and dedicated-hardware measurements.
The Missing Medium
Hytorch’s preprint introduces a transactional record of training writes, together with measured limitations.
Factual access and control design
A bilingual report preserves the V5 no-control-match result and defines the next, unexecuted V6 protocol.
Work with us
Bring a question worth investigating.
We welcome research partners, teams exploring a pilot, and sponsors who want to support open infrastructure with a public record of progress.
[email protected]ParticipateInstitution in formation
Building the conditions for lasting work.
We are forming governance, contribution, and stewardship structures around the projects. Celiums Solutions LLC is the interim organizer and custodian.
Governance