Reflective Memory

What it is

A reflective-memory AI agent built on Google Cloud. On top of Vertex AI Memory Bank, it runs a nightly Firestore “dream” that consolidates the day’s raw memories into durable, higher-level ones — so the agent gets sharper over time instead of just accumulating transcripts.

The problem

Most “memory” in agents is really just a growing log. It gets long, noisy, and expensive to read back. Humans don’t keep every transcript — we sleep, and consolidate. This project asks: what if an agent did the same?

How it works

Capture — during the day, interactions land in Memory Bank as raw memories.
Dream — a scheduled job replays recent memories, clusters and summarises them, and writes back consolidated memories (and prunes the noise) in Firestore.
Recall — at conversation time, the agent retrieves the consolidated layer first, falling back to raw memories only when it needs detail.

day:    interactions ──▶ raw memories (Memory Bank)
night:  raw memories ──▶ [ dream: cluster · summarise · prune ] ──▶ consolidated (Firestore)
recall: consolidated first ─▶ raw on demand

What I’d highlight

Treating memory consolidation as a background, scheduled process, not an inline cost.
A two-tier recall strategy that keeps prompts small and cheap.
Clean separation between the raw store (Memory Bank) and the consolidated store (Firestore).

Stack

Python · Vertex AI Memory Bank · Firestore · Gemini · ADK

Companion: Reflective Memory Skill — packaged as a reusable agent skill.
Public memory demo: Memory Agent
Concept: Agent memory

Screenshots and a walkthrough video go here.