Hippocampus

A vector database built for AI agents — not retrofitted from document search, but designed from the ground up for how agents actually manage memory.

Inspiration

The hippocampus is the part of the brain responsible for forming and retrieving memories. AI agents are becoming more powerful every day, but they remain unreliable without memory.

We built Hippocampus — a database that gives agents persistent, semantic memory so they never forget what matters.

What it does

Hippocampus is the core nervous system for AI agents, giving them persistent, semantic memory so they can reason, recall, and act consistently.

It is a vector database built specifically for AI agents, supporting two interaction patterns:

Agent-Controlled Memory – Agents manage their own memories directly.
Database-Curated Memory – An internal curation agent decomposes text into structured, meaningful memories.

The web UI is purely for non-technical users; AI agents and developers interact with Hippocampus directly via APIs

Key Features

Per-agent isolation
100ms retrieval
Agent-to-agent orchestration
Production-ready, fully serverless architecture on AWS

Real-world impact: In our safety demo, an agent remembered a child’s shellfish allergy weeks later — preventing a dangerous mistake when a parent tried to buy shrimp.

How we built it

Custom 512-dimensional indexing using sorted arrays and binary search
Binary serialization with file-based storage (inspired by SQLite)
AWS stack: Lambda (pure functions), EFS (per-agent files), S3 (backups), and Bedrock (for embeddings and reasoning)
Agent orchestration: External agents can call internal curation agents
Built with Go (backend), Python (demos), and Terraform (infrastructure)

Challenges we ran into

Cross-region Bedrock model availability — solved by making the region configurable
Overly broad search results — solved with threshold and top-k filtering
Lambda timeouts with large memories — delays are now configurable
AWS SDK v1/v2 conflicts — fixed by generating configurations dynamically
Scalability — redesigned from EC2 to pure Lambda for seamless scaling

Accomplishments that we're proud of

Built a vector database from scratch, not on top of existing systems like Pinecone.
Implemented real infrastructure: indexing, serialization, and search algorithms from first principles.
Beat FAISS by a significant margin in the specific tests that agents actually need — exact search on small to mid-size memories, optimized for latency and determinism rather than throughput.
Achieved autonomous agent-to-agent orchestration.
Delivered a real-world safety demo with measurable impact.
Fully serverless, multi-tenant, and scales elastically with Lambda.

What we learned

Simplicity scales. Binary search is fast, deterministic, and reliable — billion-scale complexity isn’t needed for agents.
Agents need databases designed for them, not repurposed document stores.
Serialization is challenging but powerful when done right.
Serverless architectures are a perfect match for agent workloads when designed around stateless, file-backed operations.

What's next for Hippocampus

Near term:

Integration with Bedrock Agents
Observability dashboard
Batch operations

Long term:

A fully managed AWS service for agent memory
Deep integrations with Bedrock Agents, Amazon Q, and Amazon Connect
Make Hippocampus the SQLite of AI agents — lightweight, open source, and ubiquitous.