Acknowledged means durable
Nexir answers OK only after the write is on disk on a majority of nodes. A three-node cluster keeps committing with one node down, and a write you were told about survives the loss of the node you wrote to.
Public beta · 0.1.0-beta.5
State your agents can't lose.
Nexir is durable, replicated state for AI agents and the systems around them: every acknowledged write is on a majority of nodes before your code moves on, reads are linearizable by default, and failures are named, never silent. Redis clients, no new SDK.
curl --proto '=https' --tlsv1.2 -sSf \
https://nexir.com/install.sh | sh -s -- --accept-terms
Durable
Acknowledged = on a majority
Linearizable
By default; local when you choose
RESP2
Redis wire protocol and clients
Public beta: evaluate with replaceable data. Interfaces, behaviour and the storage format may still change before general availability. What changed in 0.1.0-beta.5 · what Nexir does not do yet.
Built for state that matters
Three promises.
Each one checkable.
An agent's checkpoint, lease or run state is only useful if the next step can trust it. Nexir makes three commitments about every write and read, and each is backed by code and by measured runs rather than by a slogan.
Linearizable by default, local when you choose
Every read returns the latest committed value unless a connection asks for local reads, which are faster and may lag a little. The choice is explicit and per connection, so you never guess which you got.
Nothing fails silently
When Nexir cannot serve a request it says so by name — the leader moved, the cluster has no quorum, the node is overloaded — and a request that gets no reply is documented as unknown, never assumed lost or assumed done.
Beta.5, measured
Steady under large, sustained workloads.
Beta.5 makes memory use more predictable, cleans up overwritten data as part of normal storage maintenance, and helps a replica that falls behind catch up without starting over. The results below come from named runs, each recorded with its conditions; they are evidence, not capacity guarantees.
Read the release notesSee the benchmarkDemonstrated in 0.1.0-beta.5
- 1 billion writes at 51,551 per second and 2 billion at 52,193 per second on three NVMe nodes with full TLS — one run each, every key verified after a full restart
- Disk use after one billion writes about 70% lower than the same test on an earlier build without the new cleanup, and a second billion overwrites added little
- Resident memory stayed below half of each node's memory target in both runs, with no out-of-memory event
- A replica that falls behind can catch up by replaying recent writes within bounded limits, instead of receiving the whole dataset again
- A node that finds a new snapshot would not fit on disk waits and retries before building it, instead of stopping
- Leader loss, failover and rejoin are exercised on the release binary as part of release validation
Honest scope
What Nexir does not do yet.
And where it is going.
Said here so nobody has to discover it in production. The full list, with what to do instead, is in the limitations.
Not in beta.5
- No password or ACL — mutual TLS is the only client authentication
- No key iteration (
SCAN,KEYS) and no hashes, lists, sets, streams or transactions - No backup, restore, export or import tooling
- No data migration between storage-format generations, no rolling upgrade
- No hard memory ceiling and no guaranteed recovery time for a replica that falls behind
- Measurements are single-region, single-zone
Planned next — no dates promised
- Listing the keys of a run or session (prefix scan)
- Version-conditioned writes, so concurrent agents cannot overwrite each other unnoticed
- Token-based client authentication alongside mTLS
- Shorter rejoin times at large dataset sizes, and a disk-capacity governor
- Moving leadership while writes continue
Each item is tracked in the engineering roadmap with the evidence behind it; progress is reported in the release notes.
Simple deployment
One node to start. Three when you need quorum.
Run Nexir on a laptop, VM, bare-metal host or Docker. Start with a single node, then add learners and promote voting members through explicit, audited cluster commands.
Deployment guideWhat Nexir provides today
- Scalar keys and values:
GET,SETwith conditions and expiry,DEL, counters, the TTL family - Single-node and explicitly managed multi-node clusters with leader routing
- Catch-up for nodes that fell behind: replay of recent history within bounded limits, or a snapshot validated before it is accepted
- A memory plan sized from the host or container, with clear retry replies under memory pressure
- Client and peer mutual TLS, in separate trust domains
- Prometheus metrics for Raft, storage, snapshots, garbage collection, disk, memory, replica catch-up and TLS
- Bounded shutdown with a meaningful exit status, and offline majority-loss recovery
Start where you are
Linux, macOS, and Docker.
Linux
x86_64 and ARM64 binaries.
macOS
Intel and Apple Silicon binaries.
Docker
Persistent, non-root container workflow.
Manual binaries
Archives and SHA-256 checksums.
Versioned installs
Pin or upgrade a beta release.
Open source component
Nexir MVCC
Nexir is built on nexir-mvcc, a separate, deterministic MVCC library that remains open source under MIT or Apache-2.0.
Use it independently
Build transactional versioning into another storage system by implementing its backend trait. The library owns MVCC semantics — not networking, Raft, server storage integration, or the complete Nexir server.