| name | always-on-memory-agent |
| description | Use this skill when evaluating agent memory options, when running or reviewing the always-on memory agent sample (Google generative-ai repo, by Shubham Saboo), or when deciding between a self-hosted memory loop and the managed Vertex AI Memory Bank. Covers how the sample works and its rough edges, and a verified Memory Bank quickstart. SDKs used, Google ADK, google-genai, vertexai agent engines, Gemini 3.1 Flash Lite. |
Always-On Memory Agent Skill
A tested comparison, not a build. It covers an existing always-on memory agent sample and the managed Vertex AI Memory Bank, so your agent can help you run the sample, expect its rough edges, and pick the right option.
[!IMPORTANT]
The sample is not ours. It lives in Google's repo at github.com/GoogleCloudPlatform/generative-ai/tree/main/gemini/agents/always-on-memory-agent, written by Shubham Saboo. This skill tests it and compares it, it does not reimplement it.
[!IMPORTANT]
Model id gemini-3.1-flash-lite, GA, not new (announced March 2026, GA on Vertex May 2026). Vertex AI Memory Bank is GA. Neither is a new capability, the point is the honest comparison.
The sample in one look
- Four ADK
Agents, an orchestrator routes to ingest, consolidate, query. Tool calling, not output_schema.
- Always-on, an aiohttp API on :8888, an inbox folder watcher, and a consolidation loop every 30 minutes.
- SQLite store with summary, entities, topics, connections, importance per memory.
Run it (from the sample repo):
pip install -r requirements.txt
export GOOGLE_API_KEY=your_ai_studio_key
python agent.py
curl -X POST localhost:8888/ingest -d '{"text":"Sascha streams Nordlicht Friday"}'
curl "localhost:8888/query?q=when+is+the+stream"
Rough edges, measured when running it
- Retrieval is a recent-window load, not search.
read_all_memories is SELECT * FROM memories ORDER BY created_at DESC LIMIT 50. There is no vector search in the sample. Past 50 memories, older ones silently drop out of the query context.
[!WARNING]
Symptom, a query stops finding an older fact once the store grows past 50 memories. Cause, the LIMIT 50 recent-window load, there is no similarity retrieval to pull it back.
- Consolidation is capped at LIMIT 10.
read_unconsolidated_memories reads at most 10, so a burst of more than ten new memories only partly consolidates per pass.
- Per ingest latency 2.4 to 4.7s. Every ingest routes orchestrator to sub agent to tool and back, measured on 3 files.
- One shared SQLite file, no per user isolation. Fine for a personal assistant, not multi user.
- Thin governance. A delete endpoint exists, no retention, audit, or scoped access. The repo frames it as a reference implementation.
The managed alternative, Vertex AI Memory Bank
Same three inputs, run for real during the episode. Memory Bank merged them into fewer, user-scoped memories and retrieved by vector similarity search in 0.7s, versus the sample loading its recent window.
[!IMPORTANT]
Memory Bank retrieval is a scoped vector similarity search, and memories are stored first person per user_id. This is the core difference from the sample, the sample removed the vector search, Memory Bank keeps it.
scripts/memory_bank_quickstart.py, verified, point it at your own file and question.
gcloud auth application-default login
python scripts/memory_bank_quickstart.py --project YOUR_PROJECT --file notes.txt \
--query "what do you know about the project?" --user alice
The real calls are client.agent_engines.create() (engine create was 4.3s in prep), generate_memories(direct_contents_source={"events": [...]}, scope=...), and retrieve_memories(similarity_search_params={"search_query": ...}).
When to use which
- Learn the mechanics, run something small and personal, or stay local and offline, the sample is a great teaching tool and it works.
- Real users, isolation, retention, and retrieval that scales past a recent window, use managed Memory Bank.
Supporting files
- scripts/memory_bank_quickstart.py, the verified Memory Bank create, generate, retrieve utility, parameterized.
- reference.md, the full verified head to head, both run on the same inputs.
- sample-files/, example multimodal inputs (a note, a chat screenshot, a voice memo) to feed either system.
Dependencies and Prerequisites
- Python 3.12 recommended.
google-adk >= 2.4.0, google-genai >= 2.12.1 for the sample, google-cloud-aiplatform >= 1.161.0 for the Memory Bank quickstart.
[!WARNING]
The macOS system python is 3.9 and both ADK and google-genai require >=3.10. On 3.9 pip serves a stale ADK 1.x and the API looks broken. Use uv venv --python 3.12.
Documentation Pages
You MUST fetch the matching page before writing code. These hosted docs are the source of truth for parameters, types, and edge cases.
From the episode
Video walkthrough, link to follow once the stream is up.
