Catalyst

FAQ

Answers about Catalyst, Meraki, AI agent memory, continual learning, judgment profiles, privacy and product availability.

Basics#

What is Catalyst?#

Catalyst is an independent AI product and research lab building systems that help artificial intelligence retain useful context, understand human judgment, learn from feedback and improve through repeated use.

Is Catalyst an AI company, AI lab or product studio?#

Catalyst operates as an independent AI lab and product studio. It researches a specific problem—how AI systems can develop continuity and improve through experience—then turns that research into infrastructure, products and public experiments.

Who founded Catalyst?#

Catalyst was founded by Pratham, a seventeen-year-old builder with a background in video editing, creative strategy, AI systems and agent infrastructure.

What problem is Catalyst trying to solve?#

Modern models can generate strong outputs but often fail to carry meaningful learning across sessions. Catalyst is building the layer that turns corrections, approvals, rejections, references and outcomes into inspectable changes in future behaviour.

What does “building AI agents that iterate and improve” mean?#

It means building agents whose future actions can change because of evaluated experience. The system preserves evidence, interprets feedback, applies relevant guidance, measures the result and updates itself in a controlled way.

What is iterative AI?#

Iterative AI is an AI system designed around a repeatable learning loop: experience, evidence, interpretation, guidance, action, evaluation and governed update.

Is iterative AI the same as continual learning?#

They overlap. Continual learning is the broader machine-learning problem of learning from a stream of experience without losing useful prior knowledge. Catalyst uses iterative AI to describe product systems that apply those ideas through memory, feedback, evaluation and governed updates.

Is Catalyst building AGI?#

No. Catalyst is not training a general-purpose frontier model. It is building infrastructure around existing models so they can become more useful to specific people, teams and workflows.

Meraki#

What is Meraki?#

Meraki is Catalyst’s first major product. It turns references, approvals, rejections and corrections into a living, evidence-backed profile of judgment that can guide future AI work.

Is Meraki a model?#

No. Meraki works above models. It can provide context, preferences, boundaries, evidence and evaluations to an agent using a hosted or open-weight model.

What is Meraki Core?#

Meraki Core is the infrastructure beneath Meraki. It manages sources, observations, signals, profile atoms, guidance packs, run records, feedback, evaluations and controlled update proposals.

What is Meraki Studio?#

Meraki Studio is the control plane for inspecting and managing what the system has learned, where each belief came from, which tasks it affects and whether it is improving results.

Is Meraki a Pinterest alternative?#

The reference feed may feel familiar, but the purpose is different. Saving and rejecting material creates structured evidence for an AI judgment profile rather than only organizing visual inspiration.

Who is Meraki for?#

Meraki is being designed for creatives, founders, AI power users and teams whose work depends on context, standards, judgment and repeated feedback.

What can Meraki learn?#

Meraki can represent bounded preferences, constraints, references, anti-patterns, project-specific standards, working patterns and evaluation criteria. It should not infer sensitive or universal claims without clear evidence and permission.

Does Meraki copy my style?#

The goal is not surface imitation. Meraki is intended to understand the reasoning behind choices: what you protect, reject, prioritize and consider successful in a specific context.

Can Meraki work with Claude, ChatGPT, Codex, Cursor or other agents?#

The intended architecture is model- and tool-flexible. Meraki Core exposes APIs and an MCP-compatible connector so different agents can retrieve task-specific guidance. Exact integrations depend on the current release.

Memory and learning#

How is Catalyst different from AI memory?#

Memory preserves information. Catalyst focuses on whether preserved experience causes a useful, governed change in future behaviour.

How is Catalyst different from RAG?#

RAG retrieves relevant information for generation. Catalyst adds interpretation, judgment profiles, feedback, evaluation, update governance and the ability to test whether retrieved experience improved the result.

Is a long context window enough?#

No. A long context window increases how much information can be processed at once, but it does not decide what should persist, resolve contradictions, manage scope or evaluate whether old experience remains useful.

Why not store every conversation?#

More memory can create more noise, contradiction and privacy risk. Useful continuity requires selective writing, consolidation, retrieval, decay, revocation and deletion.

How does a correction become learning?#

The system preserves the correction as evidence, identifies what changed, proposes a scoped interpretation, applies that interpretation to a relevant future task and evaluates whether the change improved the outcome.

What is a profile atom?#

A profile atom is a small, scoped and evidence-backed claim about a preference, boundary, standard, goal or working pattern.

What is a guidance pack?#

A guidance pack is a task-specific compilation of the most relevant profile atoms, evidence and constraints supplied to an agent before or during a run.

Can the system learn the wrong thing?#

Yes. Human feedback is ambiguous, and automated interpretation can be wrong. That is why Catalyst treats learned claims as inspectable objects with evidence, confidence, scope and lifecycle controls.

Can I correct what the system believes?#

That is a core requirement. Claims should be confirmable, rescopable, revisable and revocable.

Does Catalyst use reinforcement learning?#

Catalyst may use evaluation and feedback mechanisms related to reinforcement-learning ideas, but the current architecture is primarily focused on governed external memory, retrieval, profile updates and task-level evaluation rather than continuously modifying foundation-model weights.

Taste and judgment#

What does taste mean in an AI system?#

Taste is contextual selection: the patterns behind what a person chooses, rejects, combines or preserves. It includes references, boundaries, standards and the reasons one plausible option is preferred over another.

Is taste static?#

No. Taste changes through exposure, creation, consequence and time. A useful system should represent changing and context-specific judgment rather than freeze a person into a permanent style profile.

Can taste be reduced to embeddings?#

Embeddings can represent similarity, but taste also depends on contrast, context, sequence, intent, consequence and rejection. Similarity is one signal, not the complete model.

Is this only for creative work?#

No. Judgment appears in coding, writing, research, operations, product decisions, hiring, analysis and any workflow where several technically valid options have different quality or fit.

Open source and models#

Is Meraki Core open source?#

Meraki Core is being developed publicly in the Catalyst GitHub organization. The repository documentation should distinguish clearly between implemented, experimental and planned components.

Is Catalyst open source?#

Parts of Catalyst are public and open source. Hosted services, managed implementations and future product layers may contain proprietary components.

What is the difference between open source AI and open-weight AI?#

Open-weight models make trained parameters available. Open source AI, under the Open Source Initiative definition, also requires the freedoms and preferred form needed to use, study, modify and share the complete system.

Does Catalyst prefer open or closed models?#

Catalyst is model-flexible. Hosted models are useful for frontier capability and operational convenience. Open-weight systems are useful for control, customization, privacy and independent deployment.

Will models become commodities?#

Frontier research will remain difficult and valuable. The commodity argument is that capable model access will become broadly available, moving product differentiation toward context, workflow integration, reliability, evaluation and accumulated user understanding.

Privacy and control#

Who owns the profile created by Meraki?#

The intended principle is that the person or organization represented by the profile should control it and be able to inspect, export and delete it.

Does Catalyst train foundation models on user data?#

Catalyst is not currently training a foundation model on user data. Any hosted processing, retention or model-provider use must be disclosed in the relevant product privacy documentation.

Can profile data be exported?#

Exportability is part of the architecture. The exact export format and availability depend on the current product release.

Can a source be deleted?#

Meraki Core is designed around deletion lineage so that removing a source can propagate to dependent observations and profile claims where required.

Can learned preferences be revoked?#

Yes. Revocation is a core concept. A revoked claim should stop influencing future guidance while preserving the appropriate audit history.

Is Meraki local or hosted?#

The long-term architecture supports a hosted canonical profile with connectors into different tools, while parts of the core can also be run locally. Current availability must be stated on the product page.

How does Catalyst handle sensitive information?#

The system should minimize collection, preserve provenance, apply scoped access controls and avoid inferring sensitive traits without explicit need and permission. Production policies will be documented before hosted release.

Product status#

Is Catalyst available now?#

The Catalyst site, public writing and Meraki Core repository are available. Meraki product access depends on the current early-access release.

Is Meraki available now?#

Meraki is in development. The landing page should state whether the current experience is a public demo, waitlist, private alpha or open release.

Is there a waitlist?#

A waitlist will collect early users and teams whose workflows depend on repeated AI judgment and feedback.

How much will Meraki cost?#

Pricing has not been finalized. Early versions may include a free entry point, paid personal features and managed or team implementations.

Can Catalyst build a system for my team?#

Managed calibration and implementation work may be available for selected founders, creatives and AI-native teams. Contact Pratham directly by email.

How do I contact Catalyst?#

Use the Contact link, which opens an email to pratham@itscatalyst.com.

Where is the code?#

Public repositories are available at the Catalyst GitHub organization: https://github.com/itscatalyst.

Research and limitations#

Is Catalyst academic research?#

Catalyst is an independent product and research effort, not a university laboratory. Claims should be grounded in experiments, public research and reproducible evaluation where possible.

What are the main technical challenges?#

The hardest problems include interpreting ambiguous feedback, avoiding negative transfer, selecting the right memory, managing contradictions, preventing profile drift, evaluating causal improvement and protecting privacy.

Can continual learning cause forgetting?#

Yes. New experience can interfere with previously useful knowledge. External memory changes the form of the problem but does not remove it; representation and retrieval can still create negative transfer.

How will Catalyst prove that learning works?#

Through held-out tasks, before-and-after comparisons, correction-rate reduction, retrieval tests, component ablations and evidence that removing a learned component removes the claimed improvement.

What is not built yet?#

The public documentation must maintain an explicit limitations page covering incomplete hosted infrastructure, authentication, production multi-tenancy, model adapters, large-scale evaluation and any simulated demo behaviour.

Agent architecture and context engineering#

What is an AI agent?#

An AI agent is a system that uses a model together with instructions, tools, state and a control loop to pursue a task across one or more steps. The model supplies general capability; the surrounding system decides what it can access, remember, verify and do.

What is agentic AI?#

Agentic AI refers to systems that do more than answer a prompt. They can plan, call tools, inspect results, update state and continue toward an outcome within defined permissions.

What is context engineering?#

Context engineering is the design of the information supplied to a model at each step: instructions, retrieved evidence, working state, tool results, examples and constraints. A larger context window does not remove the need to select and organize that information well.

How is context engineering different from prompt engineering?#

Prompt engineering focuses on the wording of an instruction. Context engineering focuses on the complete information environment around the instruction, including retrieval, memory, state, tools, examples, policies and compaction.

What is provenance?#

Provenance is the trace connecting a memory, preference or rule back to the evidence and process that produced it. Provenance makes it possible to inspect why guidance exists and to revise or delete it safely.

How does Meraki handle contradictory preferences?#

The intended system keeps contradictory evidence rather than flattening it immediately. Preferences can be scoped by project, task, audience and time, allowing the system to represent several valid modes instead of forcing a single universal personality.

How does Meraki decide what should persist?#

Durable updates should require evidence, scope and evaluation. A passing request should not automatically become a permanent preference. Meraki Core is designed to distinguish raw observations from confirmed profile claims.

Does Meraki fine-tune the underlying model?#

Not in the current public direction. Meraki primarily changes the context and guidance supplied to existing models. Future research may compare external memory and in-weight learning, but no weight-training capability should be claimed unless it is implemented and evaluated.

What are update proposals?#

Update proposals are suggested changes to a judgment profile. They allow the system or user to inspect what might change before the change becomes durable.

How are learned updates evaluated?#

A learned update should be tested on relevant held-out tasks, adjacent tasks and cases where it should abstain. The goal is to verify improvement without causing unrelated regressions or false preference application.

Can individuals and teams have different profiles?#

That is part of the intended model. Personal judgment, team standards and project-specific constraints should remain separate but composable, with explicit precedence and provenance.

Does Meraki support MCP or APIs?#

Meraki Core is being designed with API and MCP-compatible access so agents can request task-specific guidance and return feedback. The public documentation must state exactly which interfaces are implemented in the current release.

Can Meraki work with local models?#

The architecture is intended to remain model-flexible. Local and open-weight models should be usable where the connector and deployment environment support them.

Can Meraki run offline?#

Parts of the core may run locally, but complete offline support depends on the active model, storage and connectors. The public page must not claim full offline operation unless it has been tested.

How does Meraki protect against prompt injection?#

Memory and retrieved content must be treated as untrusted data, not automatic instructions. Production implementations require content boundaries, permission checks, tool policies, sanitization and adversarial evaluation.

Is Meraki a self-improving AI system?#

Meraki explores a constrained form of improvement through external evidence, feedback and governed updates. It should not be described as autonomous self-improvement in the broad AGI sense.

Access, collaboration and roadmap#

How do I join the Meraki waitlist?#

Use the waitlist form on the Meraki page. The form should clearly state the current product stage and what early users may be contacted about.

Can I contribute to Meraki Core?#

Public contribution rules depend on the repository's current status. The GitHub README should state whether issues, discussions or pull requests are open and which areas are suitable for contribution.

Is Catalyst hiring?#

Catalyst is currently an independent lab. Any open collaboration, internship or paid role should be listed publicly rather than implied through the FAQ.

Can researchers collaborate with Catalyst?#

Yes. Researchers and builders working on agent memory, context, evaluation, continual learning, human feedback or judgment systems can contact Pratham by email with a specific idea or experiment.

Can founders or teams test Meraki?#

Selected early users may be invited to test the product or a research prototype. The page must distinguish a demo, waitlist, private alpha and generally available product.

Can someone invest in Catalyst?#

Investment conversations can begin through direct email. The site should not claim that Catalyst is fundraising unless a fundraising process is actually active.

Where is Catalyst based?#

Catalyst is internet-native and currently founded from India. Location should not be represented as a registered office or corporate presence unless that is legally accurate.

What is the Catalyst roadmap?#

The roadmap is to prove the smallest learning loop, evaluate it, expose it through Meraki Core, build an inspectable product layer and expand only where real user evidence supports the next step.

What should I read first?#

Start with “AI Should Learn From You,” then read the Iterative AI documentation, “Memory Is Not Learning,” and the Meraki Core overview.