Memory

Search indexes your documents. Memoth remembers them.

A vector store makes text findable. Memoth reads each source and keeps what a person would have: what happened, what's true, how it's done, each as a typed fragment that names its origin.

The Memoth platform

Three things your agents need from memory. One layer, under all of them.

The memories a person would keep.

Point Memoth at your docs, tickets, chats and meeting notes. It keeps what a person would, typed and tagged, with the source's own permissions on every fragment.

Confluence · Payments space · 1,284 fragments extractedsynced 4 min ago
TypeMemoryAudience
SEMANTICPayments runs active-active across ap-south-1 and ap-southeast-1.devops, sre
PROCEDURALFailover runbook: freeze deploys → drain → flip DNS → verify → unfreeze.devops
EPISODIC14 Aug failover: DNS TTL 300s delayed cutover by 9 minutes.devops, sre

A ledger you can read.

Memory stays in your environment. Every merge, update and drop is recorded, every read leaves a receipt, and you can replay what an agent was served on any date.

Ledger · payments domainreplay as of 20 Aug 2026 ▾
WhenEventFragmentActor and receipt
09:14:02served 3 of 41 candidatestopology · runbook · incident #241devops-agent · #4821
09:13:40merged write-back into proceduralretry window 30s → failover runbookdevops-agent · #4822
09:13:40withheld 1 fragmentFY27 pricing · audience financepolicy · #4821

Only what fits. Or nothing.

Each agent gets a working memory sized to its persona, domain and task. When nothing fits, it gets nothing, and says so.

One store · two agents · the same question: "what changed in payments this month?"
devops-agent · persona: on-call · domain: payments
EPISODIC
14 Aug failover, 9-minute delay from DNS TTL.
PROCEDURAL
Runbook updated: retry window now 30s.
WITHHELD
FY27 pricing · audience finance
support-agent · persona: tier-1 · domain: billing
SEMANTIC
Refund window on annual plans is 30 days.
WITHHELD
Incident #241 · audience devops, sre
ABSTAIN
no permitted memory on "settlement delays" · agent told to say so

Connectors

Reads from 18 sources.

NotionGitHubGranolaDocument uploadConfluenceJiraSlackGoogle DriveGoogle DocsLinearHubSpotClickUpFirefliesZimeSharePointBitbucketMicrosoft TeamsAsana

Measured, not promised

We publish the numbers behind every claim.

Recall@30 · LoCoMo
78.4%
vs 59.9% plain text search
LLM calls · memory layer
0
deterministic, auditable
Task success · behavioural
40% → 80%
hallucination cut to a third
Promotion benchmark
CAMP-Bench
every release is judged on it

These figures come from our own evaluation suites: reproducible on a frozen fixture, not third-party validated. We run the same suites on your corpus during a pilot.

Design partners

Give your agents a memory they're allowed to have.

We're building with a handful of design partners. We measure token cost, task success and served-stale rate on your corpus, in your environment, and hand you the numbers. If they don't show up, you walk with a free audit.