The harness, tuned and deployed for you.
The harness is where MemRails runs inside your agent loop — the retrieval stack, the packet contract, the bindings to the runtimes you already use. Wire it yourself with the open library, or skip the setup entirely: a pre-tuned OpenClaw harness, calibrated for the contract and deployed in one click.
Run it yourself, or let us run the best one.
Same protocol underneath, two ways to operate it. The open library drops into your harness and writes to plain Markdown. The managed tier is a pre-tuned OpenClaw deployment we operate end to end — the SOTA harness experience, without the tuning.
You wire it into your harness.
Install the library, point it at your knowledge directory, and call memory.query() before any model call. No hosted dependency, no graph to maintain — it runs in your process and versions to Git.
- +Runs in-process, inside your existing harness
- +You bring the model key and tune the stack yourself
- +Markdown-canonical, Git-versioned, no lock-in
- +Free — orchestration billed per packet, nothing for the harness
We deploy the best one for you.
A managed OpenClaw harness, pre-tuned for the packet contract and deployed on our infrastructure in one click. The retrieval stack is calibrated, Compress-v1 is bound, the integrations are pre-wired. You get the SOTA experience with no setup.
- +Pre-tuned OpenClaw — thresholds and stack calibrated for you
- +One-click deploy to a live endpoint — no infra to run
- +Compress-v1 bound and managed; integrations pre-connected
- +Eject to self-host any time — the config is yours
The stack arrives already calibrated.
A harness is only as good as its tuning. The managed OpenClaw ships with every layer of the retrieval stack wired and the dials already set — the work most teams spend weeks getting right is done before your first query.
The L1–L5 stack, wired.
Literal scan, key lookup, semantic rank, evidence filter, and compression are connected and ordered out of the box. Cheap filters run first; the model is the last resort, not the default.
Compress-v1, bound.
The compression layer points at our managed model, fine-tuned for the packet contract. No key to provision, no endpoint to wire — packets come back with provenance and contradictions intact.
Confidence, calibrated.
Evidence thresholds, citation density, and contradiction handling are tuned against real network traffic — not left at defaults for you to discover the hard way in production.
# calibrated against live network traffic — do not hand-edit retrieval: [grep, key, semantic, evidence, compress] compress: model: compress-v1 # managed, no key required max_tokens: 600 fidelity_floor: 0.85 evidence: min_confidence: 0.75 citation_density: balanced integrations: auto # pre-wired to your harness
From corpus to live endpoint in one step.
Point the managed harness at your knowledge directory and deploy. Provisioning, tuning, and binding happen behind one action; what you get back is a live endpoint your agents can query immediately.
$ memrails harness deploy --managed ◎ provision openclaw ok 1.8s ◎ index knowledge/ ok 412 keys ◎ apply pre-tuned config ok ◎ bind compress-v1 ok ◎ wire integrations ok 9 runtimes ◇ endpoint https://hx.memrails.dev/agent-7f3a ◇ status live · querying enabled
Managed convenience, never managed lock-in.
The premium tier buys you setup and tuning, not a cage. The harness writes to the same Markdown, the config is yours to read and keep, and you can eject to self-host the day it stops earning its place.
Markdown underneath, always.
Managed or not, memory is plain Markdown with typed frontmatter, versioned in Git. The harness operates on it; it never becomes a proprietary store you can't read.
Take the tuning with you.
The pre-tuned configuration is yours to export. Eject to the open library and the same calibration carries over — you keep the weeks of tuning you didn't have to do.
The packet doesn't change.
Self-host and managed emit the identical packet contract. Switching tiers never touches your agent code — the interface it depends on stays stable.
We keep you by being good.
Nothing of yours is trapped inside MemRails, so the only reason to stay managed is that the harness is genuinely better to run than the one you'd build. That's the deal.
Scoped to your stack before you sign.
Bring your corpus, your runtimes, and your projected volume. We'll scope a managed OpenClaw deployment — pre-tuned config, deploy plan, and an SLA in writing. Or start free with the open library today.