A new discipline for agents that persist. kulla provides memory, behavior modification, and configuration, for all agents in all environments.
Everything follows from one change: move the agent's memory out of the chat window and into a record it owns. Four properties fall out of that discipline.
If you're a Chief AI Officer, VP of AI, or Head of Applied AI accountable for the autonomous-agent projects coming online, you've hit this wall. If you're a Chief Data Officer or CISO responsible for the audit trail and permission gateway around those agents, you know today's tools don't cover the gap. kulla is built for both.
On any context wipe, model upgrade, or Monday-to-Thursday gap, the agent re-reads its record and resumes exactly where it left off, with no re-onboarding tax. Projects, clients, decisions, and rationale hold across the whole engagement, even as people rotate off.
Selective context retrieval means agents pull only the sections they need, so they never blow the cache — sustaining a 99.995% median cache-hit rate on real workloads. Shared context is cached once and read many, and a mid-tier model on kulla often matches a frontier model running stateless. Net: roughly 75% lower inference cost at enterprise scale (modeled).
The encrypted memory lives on infrastructure you own, never a vendor's servers, never their training set. Move Anthropic → OpenAI → Google on a Tuesday and every decision travels with you. The model is negotiable; the discipline is the identity.
Every decision and action is written as it happens: attributable, timestamped, and (in enterprise) hash-chained. Every agent action is gated by an operator-set level: deny · ask · notify · auto. Prompt injection is closed at the framework layer, not patched at the model.
kulla is the framework. kulla Studio is the app that runs it. kulla Enterprise governs a federation of agent workspaces, and kulla Local, the engine it governs, ships inside it. Buy Enterprise and you get both. Local also stands alone for a single operator.
Everything Local has, plus a governance layer that makes agentic AI safe to deploy across an organization at scale.
The full framework, on one machine, through a GUI. Your memory in your encrypted database, your model keys, your rules.
Better models ship every month. kulla is the discipline that turns them into colleagues who hold across weeks, not tools that reset every session.
kulla is not a memory layer, an agent framework, or an observability tool alone. Memory-only tools store what an agent has learned. Agent frameworks help you build agents. Observability tools watch what happens. kulla is the memory-and-behavior layer that ties all three together: persistent context, enforced behavioral discipline, and decision continuity across sessions and across agents, deployed on your infrastructure so your data never leaves it.
The saving isn't one number. Every workload below was run locally through the full Cheapifier pipeline with the real model, and the reduction depends on the kind of work and on how much of it qualifies. We publish both — measured, not projected.
Tool-intensive automation reaches 89%; broad, mixed knowledge work lands near 58%. Each figure is that scenario's actual sequential token total through the pipeline, reproduced exactly by the calculator.
Only the tools a task needs are exposed. Up to ~83% measured when a broad catalog would otherwise be sent eagerly — the dominant structural saving.
Each task runs on the smallest capable model. ~51% measured cost cut versus always choosing the strongest, with every capability, risk, and context-window floor satisfied.
Identical reads are served from cache; mutating calls always execute. ~47% measured on a repeat-heavy fixture.
The working context is fit to a priority budget before dispatch. ~37% measured under history pressure, 100% of required items retained, overflow fails closed.
Meaning-preserving compression on what remains — 100% of protected probes retained. Modest by design (~7% end-to-end), applied last, never before the structural reductions.
The Cheapifier only acts on the share of work that qualifies. That share rises with AI maturity — from a pilot touching a fraction of workflows to an AI-native organization where nearly everything runs through it. Your savings scale with both the work mix and this coverage.
Observed, not guaranteed. 3,447 of 3,447 automated checks and 867 of 867 semantic probes passed across 93 local cases; a separate independent-agent audit re-extracted meaning from every modified context and lost none. The model, tokenizer, and report hashes are pinned. Want a number for your own seats, work mix, and provider? Build an evidence-bounded operating case →
Six disciplines kulla Studio enforces by construction, automatic in the software, not folklore your engineers pass around.
Every decision, action, and claim is captured the moment it occurs: nothing summarized, actually recorded. Export it to a regulator, query it live, or hand it to a client.
Agents persist across sessions, model upgrades, and reboots. Pause one Monday, resume it Thursday: same identity, same context, same next step.
Decision-before-code. Plan before multi-day work. Every claim carries what was checked and what was found. The app blocks non-compliant moves: a constraint, not a convention.
Every record carries what was asserted, how it was checked, and what was observed. It's impossible to record a verification that never ran: no agent can claim a check it didn't perform.
First-party connectors for secrets managers, databases, and metrics, installed from the app itself. Extend the workbench without leaving the workflow.
kulla Enterprise adds an admin surface on top: a shared constitution, a fair resource broker, cross-department countersignature policies, and hash-chained audit export.
MOMUS DEV builds kulla and stands behind it, from a single laptop to the enterprise. Founded and led by Travis Winegar, with founding partner William Chapin (COO track). kulla started from a simple frustration: agents that were brilliant for an hour but useless in three. The fix was giving the model an external memory worth keeping.
We believe all software should be secure, safe, and auditable, and that AI, of all software, is where those three stop being nice-to-haves. Durable AI is an operating commitment, not a download.
All kulla intellectual property, including framework, source code, documentation, and brand assets, is owned outright by MOMUS DEV, LLC, with clean provenance and no third-party encumbrances.
The people behind the company:
Straight answers to what buyers and engineers ask first.
kulla is the agent memory and behavioral framework for enterprise AI. It moves an agent's working memory out of the chat window into a record it owns, so agents keep their context, judgment, and decisions across sessions, model upgrades, and restarts.
Memory-only tools store what an agent has learned. Agent frameworks help you build agents. Observability tools watch what happens. kulla is the memory-and-behavior layer that ties all three together: persistent context, enforced behavioral discipline, and decision continuity across sessions and across agents.
Yes. Your encrypted agent memory lives on infrastructure you own. It never lands on a vendor's servers and never enters anyone's training set.
No. The memory lives in your files, not the provider's. You can move between Anthropic, OpenAI, Google, or a local endpoint without re-onboarding a single agent — the decisions travel with you.
Selective context retrieval means agents pull only the sections they need, so they don't blow the cache — sustaining a 99.995% median cache-hit rate on real workloads. Combined with model routing, that lands roughly 75% lower inference cost at enterprise scale (modeled).
Every decision and action is written as it happens. kulla Enterprise adds a tamper-evident, hash-chained audit ledger and a deny-by-default permission gateway where every agent action is gated as deny, ask, notify, or auto. Prompt injection is closed at the framework layer, not patched at the model.
kulla Local is the single-operator studio: the full framework on one machine with your own model keys. kulla Enterprise adds multi-department governance, a fair resource broker, countersignature validation, and a tamper-evident audit trail across the organization — and it includes the Local engine.
kulla is built and owned outright by MOMUS DEV, LLC, a software company in Missouri, USA founded by Travis Winegar, with clean IP provenance and no third-party encumbrances.
Every claim on this page is backed by measurement or by the framework itself. Tell us what your agents are trying to do, and we'll walk you through what changes when they stop forgetting: what it costs, what it saves, and what it takes to stand up.
We answer directly, usually within a business day.
Direct: sales@momusdev.com