engram MCP

I couldn't find "NexForge" specifically — is it possible you mean NexHacks? Either way, the form is clearly a standard Devpost submission. I have enough context on Engram to write all seven sections. Here they are:

Inspiration

Multi-agent AI systems are becoming the default architecture for complex tasks — but every framework treats memory as a solo problem. Each agent gets its own context, its own beliefs, its own history. The moment two agents collaborate, you get drift: conflicting facts, stale assumptions, broken coordination. We built Engram because individual agent memory is solved. Cross-agent belief consistency is not.

What it does

Engram is an MCP (Model Context Protocol) server that gives multi-agent systems a shared, conflict-aware memory layer. Agents read and write to a common memory store, and Engram automatically detects when two agents hold contradictory beliefs about the same entity. When conflicts arise, it surfaces them through a resolution dashboard, so either the system or a human operator can reconcile them before they propagate into bad decisions.

How we built it

Engram is built as a hosted MCP server with a curl-installable setup, Stripe-integrated access tiers, and a conflict resolution dashboard. The core conflict detection engine compares incoming memory writes against existing agent beliefs using semantic similarity and structured entity matching. The server exposes MCP-compatible tool endpoints so any MCP-enabled agent framework can plug in without modification.

Challenges we ran into

The hardest problem was defining what "conflict" actually means in a multi-agent context. Two agents disagreeing isn't always wrong — sometimes they've observed different things. We had to distinguish genuine contradictions from legitimate divergence, which required building a nuanced conflict taxonomy rather than a simple diff. We also had to design the memory schema to be framework-agnostic, so Engram works across Claude, GPT-based agents, and open-source pipelines without requiring custom integration.

Accomplishments that we're proud of

Getting to a live, installable product — not just a demo. Engram has a real curl installer, real hosted infrastructure, and real conflict resolution tooling. We're also proud of the positioning clarity: Engram isn't another vector store or RAG layer. It's the first memory system that treats cross-agent consistency as a first-class problem, and we've been able to articulate that distinction in a way that resonates with the research and engineering communities.

What we learned

Memory in multi-agent systems is fundamentally a distributed systems problem, not just an AI problem. The challenges — consistency, conflict resolution, eventual convergence — have analogues in databases and distributed computing that we drew from heavily. We also learned that the MCP ecosystem is moving fast, and building on top of it as infrastructure (rather than as an application) opens up significantly more surface area for adoption.

What's next for Engram MCP

Expanding the conflict resolution engine to support automatic reconciliation strategies, not just detection. Building deeper integrations with popular agent frameworks like LangGraph and CrewAI. Opening a contributor program so the community can extend Engram's memory adapters and conflict resolution policies. And continuing to push toward making Engram the default shared memory standard for production multi-agent systems.

Built With

• python