federated-agent-memory

A shared, self-learning memory layer for customer-facing AI agents. Every agent learns from every conversation — even when the customer is anonymous.

This repository is the specification store for the project. It holds the architecture, the docs, and — most importantly — the OpenSpec specs that define each component before any code is written.

Gemini Enterprise Hackathon · Track 2 (Custom Agent / MCP)

Architecture v2 (2026-08-21): we now build a smarter memory service on top of Google Memory Bank instead of a self-built storage stack. See docs/02-architecture.md and the adopt-google-memory-bank change.

How we work (OpenSpec)

We use OpenSpec — specs before code.

• **openspec/specs/** — the source of truth: how each component currently

behaves, written as requirements + scenarios.

• **openspec/changes/** — proposals: when we change a component, we create a

change folder (proposal.md + delta spec + design.md + tasks.md), and on archive its deltas merge back into specs/.

👉 New here? Read CONTRIBUTING.md for the full Spec-Driven Development workflow.

Component spec backlog (v2)

Removed from v1 (replaced by Google Memory Bank): MCP Server (L1), Raw Archive (L2), Graph Core (L4), Vector Index (L5), Metadata Store (L6), Retrieval Service (L7), Cross-Agent Enrichment (L8, folded into Ranker), Agent Injection (L9), Dashboard (L10).