GOVERN

Asking a model to keep a secret
is not access control.

Memoth never asks. Every read resolves to a database predicate: row-level security is forced on, ACLs are deny-wins, and the memory layer makes zero LLM calls. The same store and the same question return the same fragments — for that agent, and for no other.

EXTRACT · GOVERN · SERVE

What Memoth is

A store, a ledger, and a write-back loop. One layer under every agent.

MEMORY STORE
EPISODICpayments failover — what broke
SEMANTICtopology: active-active
PROCEDURALrollback runbook · 6 steps

Typed fragments, not chunks — what a mind would keep.

GOVERNANCE
source: Arch p.12 · verified
merge recorded · ledger #4821
ACL: devops → read

Every memory carries its source; every change hits a ledger.

PERSISTENT LAYER
learned: retry window 30s
written back · 2m ago
week-2 recall +18%

What agents learn writes back. Week two beats week one.

Memoth.aiThe memory layer

How it actually works

Three mechanisms, and not one of them is a prompt.

FORCED RLS

The tenant boundary is enforced by the database, not by application code that has to remember to filter. Row-level security is forced, so a query that forgets its tenant returns nothing instead of everything.

Memory lives inside your environment, partitioned by domain.

DENY-WINS ACLS

Access is data: an edge between an agent and a resource, granted and revoked like any other row. Where a grant and a deny meet, the deny wins. There is no per-fragment grant to drift out of sync — reachability is a containment test, and the product never pretends otherwise.

Revoking states its blast radius first: it removes that agent's access to every fragment from that resource.

PROVENANCE

Every fragment carries where it came from. Updates supersede rather than overwrite, and the chain reads back to the original, so a correction never silently swallows what it replaced.

Zero LLM calls inside the memory layer, so decisions are deterministic and auditable.

Measured, not promised

We publish the numbers behind every claim.

37/40
Answers correct on our document-QA suite, held to a zero-hallucination CI bar: one made-up answer fails the build.
40% → 80%
Task success with conditioned memory on CAMP-Bench, with hallucination cut to a third.
CAMP-Bench
Our benchmark for promotion: the right memory, to the right agent, at the right time, per domain. Every release is judged on it.

Both figures are reproducible on a frozen fixture, not third-party-validated. We run the same suites on your corpus during a pilot and hand you the table.

Who can read what

Access is a screen, not a config file.

Every edge below is a row an operator adds or removes, and every read is checked against it. This is the product's access surface — the names in it are examples, not a live tenant.

Access edges

product UI · illustrative
resourceagents with read accessedge
DevOps / Runbooksdevops-agentoncall-agentrevoke
Support / Refund policysupport-agentrevoke
People / Compensationno edge — nothing here is readablegrant

An agent reads a fragment because it holds an edge to the resource that fragment came from. Nothing grants one fragment at a time, and nothing is readable by default.

Design partners

We're choosing a handful of teams to build with.

What you get

The memory layer under your agents early, tuned to your domain first. We measure token cost, task success and served-stale rate, and hand you the numbers. If they don't show up, you walk with a free audit.

Founder-direct iteration. Your corpus never leaves your environment.