Claude can remember now.
Deciding what is your job.
Memory for Claude Managed Agents reached public beta on 23 April 2026. On 6 May, Anthropic introduced Dreams, a research preview for memory consolidation, which extended to Fable 5 and Sonnet 5 on 10 July. Together they mark a steady build-out of Claude’s ability to retain and make sense of information across sessions.
Should Claude remember your client’s preferences from six months ago? Only the ones you have deliberately decided are worth remembering. The spring’s releases made that a real choice rather than an accident, and the judgement calls are where we spend most of a session on this topic, not the mechanics of switching memory on.
The appeal is obvious; start there honestly
An assistant that remembers a client’s preferences, a project’s history or a team’s recurring patterns without re-briefing is straightforwardly more useful than one that does not. Every re-pasted context document and every “as I mentioned last time” is friction this removes. That is the case for it, and it is a good case.
Can remember and should remember are different questions
Memory that quietly accumulates everything said across every interaction is not obviously desirable, particularly around personal or sensitive information. Teams adopting it need a clear view of what belongs in long-term memory and what stays scoped to a session, decided deliberately rather than left to defaults. This is a governance decision wearing a feature’s clothing.
Dreams is a different thing from a memory store
Rather than storing raw facts, Dreams aims at consolidating: synthesising patterns and context over time into more useful recall than a flat log of everything ever said. It is also still a research preview, which means treating it like any young feature: lower-stakes workflows first, behaviour reviewed periodically, no assumption that day-one output is where it will settle.
Note what a chat does not do
None of this changes default chat behaviour: a normal conversation still starts fresh, as our tokens guide explains. Memory here is an opt-in capability configured for agents and API-side workflows. If a colleague believes Claude is already remembering their conversations, that is a misconception worth correcting before it becomes a data-handling assumption.
Start narrow, and write the rule down
Pick one specific, low-risk thing worth remembering about your business: a client’s stated preferences, a project’s decision log. Get that working, with an explicit written rule for what is in and out of scope, before broadening. Teams that try to make Claude remember everything about everyone from day one build a liability, not an asset.
For each candidate memory, ask: if this surfaced in an output six months from now, in front of the wrong audience, is it a convenience or an incident? Convenience: store it. Incident: session-scope it. Unsure: session-scope it and revisit.
Things people
usually ask.
By default, no; each conversation starts fresh unless you bring material back in. The memory capabilities here are opt-in features for agents and API-side workflows, configured deliberately, not something switched on in an ordinary chat.
Memory is retention: an agent keeping facts and context available across sessions. Dreams is consolidation: a research-preview feature that synthesises patterns from accumulated context, aiming at more useful recall than a raw log. Retention first, consolidation on top.
With one narrow, low-risk, high-friction thing: the context document someone re-pastes weekly, a project decision log. Prove the value and the governance on that before widening scope.