Does Amazon Kinesis Remember Between Runs? Records Stay in the Stream for Up to a Year. The Next Run Still Starts With Only Its Record.

Amazon Kinesis Data Streams remembers records longer than most streaming tools do. A stream stores each record for 24 hours by default, and that window can be extended to 8760 hours, a full 365 days. Reading a record does not delete it, so several consumers can read the same record independently, and any consumer can replay records from earlier in the retention window. That makes Kinesis a durable, ordered log. It does not make it memory for the work itself. A record holds the data a producer sent, in the order it arrived inside its shard. What a run checked, decided, and finished after reading that record was never written back to the stream, so the next run starts with a record, and nothing else.

A stream keeps records, in order, inside shards

A Kinesis data stream is a set of shards. Each shard holds an ordered sequence of data records, and each record is made of three parts: a sequence number, a partition key, and a data blob, which is an immutable sequence of bytes. The partition key decides which shard a record goes to, so records that share a key stay in order relative to each other. Records with different keys can land in different shards, where no single combined order is promised.

Producers write records into the stream, and consumers read them out. In the standard model, consumers pull records and track their own position in each shard. A consumer application keeps its checkpoint outside the stream, for example in a DynamoDB lease table when the Kinesis Client Library is used. Enhanced fan-out consumers get a dedicated read path, but the basic shape stays the same: the stream serves records, and each consumer remembers, somewhere else, how far it has read.

That design is deliberate. A slow reporting consumer does not hold back a fast alerting consumer, because neither one changes the stream for the other. A record can be read once, read ten times, or skipped and returned to later, as long as it is still inside the retention window.

Retention is a replay window, not a record of work

Extending retention from 24 hours to 7 days, 30 days, or 365 days buys time to replay. A consumer that was down for a weekend can catch up on Monday. A new consumer can be backfilled with records that arrived before it existed. An investigation can re-read the exact records a decision was based on, instead of relying on a summary written afterwards.

Replay, though, returns the starting point, not the ending point. The record that comes back on replay is byte for byte the record that was written. It does not carry a note saying which run handled it, which other records were checked alongside it, what conclusion was reached, or whether the action it triggered succeeded. Nothing in the stream changes when a consumer finishes its work, because consumers do not write results into the records they read.

The window also closes. Once a record passes the retention period, it is no longer accessible, even if no consumer ever read it. A stream set to the default keeps yesterday available and loses last week. A stream set to a year keeps a year, at extra cost for the extended retention, and still treats that stored data as raw input waiting to be processed, not as a settled account of what processing already happened.

What a scheduled agent actually needs at the start of a run

Consider a scheduled automation that consumes a Kinesis stream through n8n, Make, Zapier, or a custom agent, and acts on what it finds. At the start of a new run, the next unread record is only one part of what that run needs. It also needs to know which earlier records were already investigated, so a replay or a restart does not trigger the same action twice. It needs the decisions made on those records, so a pattern checked and cleared yesterday is not treated as brand new today. It needs the full reasoning behind those decisions, so a genuinely new record that looks similar can be separated from a settled one. And it needs that record of work to be shared, so a second workflow reading the same stream starts from the same past instead of building its own.

Kinesis supplies the records, the order inside each shard, and a generous replay window. The rest is memory, and memory belongs in a layer built to hold it.

That layer is what Vilix AI provides. Vilix AI is cloud-hosted, with zero infrastructure to run: no extra stream, shard, or checkpoint table to provision beside the existing setup. The same memory is available everywhere over MCP, so the automation consuming the stream, the agent reviewing the outcome, and every other connected tool read and write one shared memory. Vilix AI stores full conversation history, not just extracted facts, so a later run can revisit what was actually discussed and decided, in context, instead of working from a single replayed record or a short summary.

The practical split is simple. Let Kinesis do the job it is built for: ingesting records in real time, holding them in order, serving them to many consumers, and replaying them inside the retention window. At the end of each run, write the outcome into memory: what arrived, what was checked, what was decided, and what still needs attention. At the start of the next run, read that memory before acting. A replayed record then meets an agent that can check the record of work first, recognize finished work, and spend the run only on what is genuinely new.

There is no cost to testing that split. Vilix AI has a free plan that stays free forever, and Pro comes with a 7-day trial that needs no credit card. The memory stays portable throughout: export everything or delete it anytime.

Get Started for Free Free forever, no credit card.

Kinesis will keep the next record, in order, for as long as the retention window allows. The memory makes sure the next run arrives knowing everything the earlier runs learned.