Agents that remember

Every time you close a session, your agent loses everything it learned. In this workshop, you'll wire persistent memory onto a Claude agent and then use Dreaming to batch-consolidate past transcripts into structured recall. By the end of the 45 minutes, you'll have an agent that remembers across sessions and you'll know how to set this up for your own agents.

May 23, 202628mWatch on YouTube ↗

WHAT IT’S REALLY ABOUT

Building Claude agents that persist memory and improve it over time

The workshop demonstrates the default limitation of agents: separate sessions are isolated, so information shared in one session is not available in later ones.
Anthropic’s Memory Stores attach a persistent, file system-like store to sessions so agents can read and write information across sessions using familiar tools like bash and grep.
Memory Stores can be configured per boundary (user/workspace/etc.), can be read-write or read-only, and can be inspected/edited manually via CLI or Console with versioned files.
Dreaming is an asynchronous, multi-agent background job that fact-checks, consolidates, deduplicates, and reorganizes a memory store using selected past session transcripts.
The workflow culminates in using the Dream-produced output memory store for improved recall and efficiency (e.g., index files), with options to review diffs and retire older stores afterward.

IDEAS WORTH REMEMBERING

5 ideas

Without shared memory, agents cannot carry context across sessions.

The demo shows that telling an agent information in one session doesn’t help a later session unless a shared memory store is mounted, limiting real-world multi-step workflows.

Memory Stores turn multi-session work into a cumulative knowledge process.

By attaching a file system-like store to sessions, the agent can persist notes (e.g., to a sessions.md file) and later retrieve them via tools like grep, enabling reliable recall.

Treat memory boundaries and permissions as a product design choice.

You can create many stores (per user, workspace, or other scope) and set access to read-only or read-write depending on whether the agent should be allowed to update shared knowledge.

Operational visibility improves trust and debugging.

The Console shows session activity (including tool use) and provides a memory-store file viewer; Dreaming also exposes its own session, making it easier to diagnose why certain memories were created or modified.

Dreaming addresses the “memory sprawl” problem by improving memory quality over time.

As memory stores grow, Dreaming runs a background, multi-agent harness to organize, enrich, deduplicate, and check staleness/facts so the store stays useful instead of becoming a dumping ground.

WORDS WORTH SAVING

5 quotes

today we're gonna talk a little bit about the base case with agents today, uh, which is that they're isolated, and this kind of limits their usefulness in a lot of real-world workflows.

— Kevin

a memory store is a persistent file system-like store that gives, uh-- that attaches as a resource to sessions that you create, and it gives agents access to, uh-- it gives agents the ability to read and write information across sessions.

— Kevin

the actual interesting thing here is that we've actually u- mounted it as a file system because it's such a powerful interface for the model.

— Kevin

dreaming is a, is a batch process that runs, again, asynchronously.

— Kevin

we don't actually touch the input memory store at all that you create. So this is a non-destructive process.

— Kevin

Session isolation in Claude Managed AgentsMemory stores as persistent filesystem mountsRead/write vs read-only memory accessManual inspection, editing, and versioning of memory filesDreaming: asynchronous multi-agent memory distillationNon-destructive cloning to an output memory storeToken usage, caching, and cost controls

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.