Google senior AI product supervisor Shubham Saboo has turned one of many thorniest issues in agent design into an open-source engineering train: persistent reminiscence.
This week, he printed an open-source “All the time On Reminiscence Agent” on the official Google Cloud Platform Github web page below a permissive MIT License, permitting for industrial utilization.
It was constructed with Google's Agent Improvement Equipment, or ADK launched final Spring in 2025, and Gemini 3.1 Flash-Lite, a low-cost mannequin Google launched on March 3, 2026 as its quickest and most cost-efficient Gemini 3 collection mannequin.
The venture serves as a sensible reference implementation for one thing many AI groups need however few have productionized cleanly: an agent system that may ingest data repeatedly, consolidate it within the background, and retrieve it later with out counting on a standard vector database.
For enterprise builders, the discharge issues much less as a product launch than as a sign about the place agent infrastructure is headed.
The repo packages a view of long-running autonomy that’s more and more enticing for assist techniques, analysis assistants, inside copilots and workflow automation. It additionally brings governance questions into sharper focus as quickly as reminiscence stops being session-bound.
What the repo seems to do — and what it doesn’t clearly declare
The repo additionally seems to make use of a multi-agent inside structure, with specialist parts dealing with ingestion, consolidation and querying.
However the equipped supplies don’t clearly set up a broader declare that it is a shared reminiscence framework for a number of impartial brokers.
That distinction issues. ADK as a framework helps multi-agent techniques, however this particular repo is finest described as an always-on reminiscence agent, or reminiscence layer, constructed with specialist subagents and protracted storage.
Even at this narrower stage, it addresses a core infrastructure downside many groups are actively working via.
The structure favors simplicity over a standard retrieval stack
In keeping with the repository, the agent runs repeatedly, ingests information or API enter, shops structured recollections in SQLite, and performs scheduled reminiscence consolidation each half-hour by default.
A neighborhood HTTP API and Streamlit dashboard are included, and the system helps textual content, picture, audio, video and PDF ingestion. The repo frames the design with an deliberately provocative declare: “No vector database. No embeddings. Simply an LLM that reads, thinks, and writes structured reminiscence.”
That design selection is probably going to attract consideration from builders managing value and operational complexity. Conventional retrieval stacks typically require separate embedding pipelines, vector storage, indexing logic and synchronization work.
Saboo's instance as an alternative leans on the mannequin to arrange and replace reminiscence instantly. In follow, that may simplify prototypes and scale back infrastructure sprawl, particularly for smaller or medium-memory brokers. It additionally shifts the efficiency query from vector search overhead to mannequin latency, reminiscence compaction logic and long-run behavioral stability.
Flash-Lite offers the always-on mannequin some financial logic
That’s the place Gemini 3.1 Flash-Lite enters the story.
Google says the mannequin is constructed for high-volume developer workloads at scale and priced at $0.25 per 1 million enter tokens and $1.50 per 1 million output tokens.
The corporate additionally says Flash-Lite is 2.5 occasions sooner than Gemini 2.5 Flash in time to first token and delivers a forty five% enhance in output pace whereas sustaining comparable or higher high quality.
On Google’s printed benchmarks, the mannequin posts an Elo rating of 1432 on Enviornment.ai, 86.9% on GPQA Diamond and 76.8% on MMMU Professional. Google positions these traits as a match for high-frequency duties corresponding to translation, moderation, UI era and simulation.
These numbers assist clarify why Flash-Lite is paired with a background-memory agent. A 24/7 service that periodically re-reads, consolidates and serves reminiscence wants predictable latency and low sufficient inference value to keep away from making “all the time on” prohibitively costly.
Google’s ADK documentation reinforces the broader story. The framework is introduced as model-agnostic and deployment-agnostic, with assist for workflow brokers, multi-agent techniques, instruments, analysis and deployment targets together with Cloud Run and Vertex AI Agent Engine. That mixture makes the reminiscence agent really feel much less like a one-off demo and extra like a reference level for a broader agent runtime technique.
The enterprise debate is about governance, not simply functionality
Public response exhibits why enterprise adoption of persistent reminiscence is not going to hinge on pace or token pricing alone.
A number of responses on X highlighted precisely the considerations enterprise architects are more likely to elevate. Franck Abe referred to as Google ADK and 24/7 reminiscence consolidation “good leaps for steady agent autonomy,” however warned that an agent “dreaming” and cross-pollinating recollections within the background with out deterministic boundaries turns into “a compliance nightmare.”
ELED made a associated level, arguing that the primary value of always-on brokers is just not tokens however “drift and loops.”
These critiques go on to the operational burden of persistent techniques: who can write reminiscence, what will get merged, how retention works, when recollections are deleted, and the way groups audit what the agent realized over time?
One other response, from Iffy, challenged the repo’s “no embeddings” framing, arguing that the system nonetheless has to chunk, index and retrieve structured reminiscence, and that it might work nicely for small-context brokers however break down as soon as reminiscence shops change into a lot bigger.
That criticism is technically vital. Eradicating a vector database doesn’t take away retrieval design; it adjustments the place the complexity lives.
For builders, the tradeoff is much less about ideology than match. A lighter stack could also be enticing for low-cost, bounded-memory brokers, whereas larger-scale deployments should demand stricter retrieval controls, extra specific indexing methods and stronger lifecycle tooling.
ADK broadens the story past a single demo
Different commenters centered on developer workflow. One requested for the ADK repo and documentation and wished to know whether or not the runtime is serverless or long-running, and whether or not tool-calling and analysis hooks can be found out of the field.
Based mostly on the equipped supplies, the reply is successfully each: the memory-agent instance itself is structured like a long-running service, whereas ADK extra broadly helps a number of deployment patterns and consists of instruments and analysis capabilities.
The always-on reminiscence agent is attention-grabbing by itself, however the bigger message is that Saboo is making an attempt to make brokers really feel like deployable software program techniques reasonably than remoted prompts. In that framing, reminiscence turns into a part of the runtime layer, not simply an add-on function.
What Saboo has proven — and what he has not
What Saboo has not proven but is simply as vital as what he's printed.
The supplied supplies don’t embody a direct Flash-Lite versus Anthropic Claude Haiku benchmark for agent loops in manufacturing use.
Additionally they don’t lay out enterprise-grade compliance controls particular to this reminiscence agent, corresponding to: deterministic coverage boundaries, retention ensures, segregation guidelines or formal audit workflows.
And whereas the repo seems to make use of a number of specialist brokers internally, the supplies don’t clearly show a bigger declare about persistent reminiscence shared throughout a number of impartial brokers.
For now, the repo reads as a compelling engineering template reasonably than a whole enterprise reminiscence platform.
Why this issues now
Nonetheless, the discharge lands on the proper time. Enterprise AI groups are transferring past single-turn assistants and into techniques anticipated to recollect preferences, protect venture context and function throughout longer horizons.
Saboo's open-source reminiscence agent provides a concrete place to begin for that subsequent layer of infrastructure, and Flash-Lite offers the economics some credibility.
However the strongest takeaway from the response across the launch is that steady reminiscence might be judged on governance as a lot as functionality.
That’s the actual enterprise query behind Saboo's demo: not whether or not an agent can bear in mind, however whether or not it will probably bear in mind in ways in which keep bounded, inspectable and protected sufficient to belief in manufacturing.
