Project Context: Reliable and Secure AI Memory

1. Project summary

Here, we propose Project Context, an open research and software initiative to develop a reliable and secure AI memory system. Our goal is to enable users to preserve and manage project context across sessions, seamlessly transfer it between AI models, and collaborate within teams while maintaining control over confidential information.

As AI agents increasingly operate across multiple sessions, teams, and organizations, maintaining reliable and secure information becomes more challenging. Existing systems face several problems, including loss of context, outdated or inaccurate memories, error propagation between agents, memory poisoning, confidential information leakage, and unauthorized information sharing.

Our work aims to adress these via three subgoals. First, build an intuitive, model-agnostic product that makes persistent memory and context transfer accessible without technical expertise. Second, develop open benchmarks for reliability, security, and multi-agent collaboration. Third, investigate the causes of memory failures through causal experiments, mechanistic interpretability, and quantitative modeling.

We plan to release the product publicly, together with open-source components, reproducible benchmarks, reusable security tools, and scientific findings. Everyday AI users should be able to use the product. Researchers should be able to reproduce our results, examine security risks, and use the tools to improve other memory systems.

2. What are this project's goals? How will you achieve them?

There has been substantial progress on agent memory, long-term reasoning, and multi-agent collaboration. Using that work often still means using a research prototype or a tool that needs technical knowledge. We want to make the capabilities available to more people, while investigating the reliability and security of systems that retain and share information.

• Objective 1: Build a user-friendly AI memory product for everyone

  What should a user be able to do? Keep project context, move it between AI models and sessions, and collaborate securely without needing technical expertise. We are building for individuals, researchers, developers, and organizations, rather than for one kind of user or one model.

  The product will integrate with supported AI tools. Its interface will let people manage memory, share context, and control information permissions. Users will be able to inspect the information being retained, correct errors, and choose what to share. Source attribution will let them check where a record came from.

  Onboarding and routine use are part of the development work. People without programming experience should not need to configure memory databases or agent frameworks to get started.

• Objective 2: Establish open benchmarks for memory reliability and security

  We will build reproducible tests for preserving context and continuing tasks across sessions. Security tests will cover memory poisoning, confidential information leakage, and unauthorized information transfer.

  To test collaboration, we will simulate organizational workflows with different agent roles, clearance levels, and permissions. We will examine who receives information, whether any recipient is unauthorized, and whether the agents complete the task. That lets us measure security alongside the effect of access restrictions on useful work.

  Our evaluation scenarios will cover cross-session memory, permission hierarchies, adversarial information injection, and multi-agent collaboration. We will compare memory architectures and security baselines using task-completion accuracy, information-leakage rates, attack success, provenance preservation, and computational overhead.

  We will publish the scenarios and the evaluation framework. Others will be able to reproduce the experiments and test their own systems.

• Objective 3: Understand and improve the mechanisms behind failures

  Counting failures is only part of the research. We also want to identify the component responsible for a failure. Was it the language model itself? Or was it something in storage, compression, retrieval, synthesis, or permission management?

  We will investigate these questions with controlled interventions, information-flow analysis, and mechanistic interpretability. Our methods include neural activation analysis and causal ablations. Distinguishing the sources of failures should help us understand which parts of a system need improvement.

  We will also study connected agents rather than treating each failure in isolation. We are interested in whether statistical scaling relationships can describe how a local failure spreads through an organizational agent network with different access hierarchies.

4. How will this funding be used?

Our full request is USD 150,000 for 12 months. It covers the work of three people—a researcher, a product designer, and a marketing and outreach specialist—as well as computational and hosting costs. The work includes scientific research, product development, computational experiments, usability testing, and public outreach.

Budget

• Team compensation, three members over 12 months: USD 110,000.

• HPC/GPU resources and mechanistic experiments: USD 22,000.

• LLM API access and evaluations: USD 13,000.

• Cloud infrastructure, storage, and hosting: USD 5,000.

Total: USD 150,000.

The minimum useful grant is USD 30,000. That would fund a three-month phase with an initial memory-management prototype, preliminary comparisons across supported models, and a public technical report with evaluation code.

The rest of the funding would support the full 12-month objectives: additional model integrations, larger security experiments, mechanistic investigations, and the complete public product release.

Timeline

• Months 1–3. We will develop the core product, design the interface, implement initial model integrations, and establish evaluation infrastructure. We aim to finish with an initial working prototype and preliminary memory-reliability evaluations.

• Months 4–6. We will expand model support, run benchmarking and security experiments, and start outreach and testing with early users. The target is at least 150 reproducible benchmark scenarios and baseline comparisons.

• Months 7–9. We will develop secure memory-sharing features and use feedback to improve the product. We will also work on provenance tracking, permission auditing, and controlled experiments on failure propagation between agents.

• Months 10–12. We will finish cross-model validation and usability studies, release the public product, and publish documentation and research outputs. Those outputs will include benchmark results, evaluation code, and reusable security components.

The researcher will lead scientific investigations, benchmarking, and technical validation. The product designer will lead the interface and product development. Community engagement, onboarding, and public communication will be handled by the marketing and outreach specialist.

We will make the grant-funded research outputs, benchmarks, evaluation tools, and reusable security components public. They should remain useful even without commercial adoption of the product. Future proprietary or commercial extensions would be funded separately.

5. Who is on your team? What's your track record on similar projects?

There are three of us. Our backgrounds cover academic research, AI software development, product design, and marketing.

• Hakan Akgün — Principal Investigator / Research Lead

  Hakan is a graduate researcher at the National University of Singapore (NUS), with a background in physics. His research has involved AI, mathematical discovery, scientific computing, and biomedical technologies. He has publications in Communications Physics, ICLR 2026, and Nature Cardiovascular Research, among others.

  His current interests are automated mathematical discovery, AI for physics, mechanistic interpretability, and reliable AI reasoning. For this project, he will lead benchmark design, causal experiments, quantitative modeling, and technical validation of memory reliability and security.

  Google Scholar: scholar.google.com

  LinkedIn: linkedin.com/in/hakan-akgun-289b39203

• Mehmet Barış Cankır — Marketing Specialist / Outreach Lead

  He co-founded Orvion Digital and is its CMO. He leads go-to-market strategy, user acquisition, and brand positioning for the studio's mobile games, mobile apps, and native AI applications.

  He was previously a Product Manager at a leading consumer app company, working on a portfolio of apps with over 120 million combined users. He has a B.S. in Industrial Engineering from Boğaziçi University and brings an analytical, metrics-driven approach to growth, user onboarding, and community engagement.

  Website: orviondigital.co

• Berkay Yalçınkaya — Product Designer / Development Lead

  He co-founded Adesse Health and leads its product design and engineering. It is a bilingual AI platform used in psychotherapy practices, with a therapist web application and a patient mobile application developed for a privacy-sensitive healthcare setting. Adesse was selected as a Top 20 global finalist at the Harvard Health Systems Innovation Lab Hackathon 2026.

  He also founded Orvion Digital, which develops mobile games, consumer applications, and AI-powered products. His degree is a B.S. in Industrial Engineering with a minor in Financial Engineering from Koç University.

  He will lead product architecture, application development, interface design, and implementation of the memory-management and permission-control features for Project Context.

  Website: adessehealth.com

  Portfolio: orviondigital.co

  GitHub: github.com/BerkayYalcinkaya

Current progress

Work on the first prototype is already underway. We are using our own time and resources, without external funding, and are currently working on the core architecture and initial implementation.

We intend to keep developing the project. Funding would give us more time for research and engineering and help us turn the prototype into a publicly accessible product. It would also support the broader scientific evaluations and release of the open research tools described in this proposal.