AIYING • V1.0 REGULAR VIEW
CLAUDE COMPLETE • POST #05
Context, Memory, Instructions and Projects: What Carries Forward?
Understand what Claude can draw on—and keep important facts anchored to current evidence.
By Ravindra Dastikop
Yesterday, you explained a client engagement to Claude. Today, you open a new conversation and write: “Continue the proposal.”
What information will Claude have? Your earlier conversation? Uploaded documents? Saved preferences? The latest approved decision?
The answer depends on several features and their settings. Treating all of them as “memory” can create mistaken expectations.
1. A practical map
The following descriptions are simplified explanations. Specific behavior depends on the Claude environment and settings.
| Element | Main role | Illustrative example |
|---|---|---|
| Current context | Information available for this task | The proposal and approval record supplied now |
| Account instructions | General guidance across conversations | Use plain English and distinguish facts from suggestions |
| Project instructions | Guidance for a particular project | Follow this engagement’s review procedure |
| Memory | Carry useful information between supported interactions | Your role and recurring working preferences |
| Past-chat search | Retrieve relevant earlier conversations | Find a previous discussion about the proposal |
| Projects | Organize related work and supporting material | A workspace for one client engagement |
2. Context: what is available for the current work?
For this article, context means the information available to help Claude perform the current task: instructions, conversation material, relevant files and information retrieved through available tools.
Information existing somewhere in your account does not establish that it has been retrieved or used correctly for a particular answer.
In our fictional engagement: a previous draft used a six-week schedule. A later approval record specifies five weeks. Your current task should identify the governing record explicitly.
Use the supplied Approved_Decisions document to determine the delivery schedule. Identify the supporting section. Flag any conflict with the earlier proposal.
This provides a checkable basis for the revision.
3. Instructions: guidance about how to work
Anthropic distinguishes account-wide instructions from instructions that apply within a Project.
Account-level example: “Use plain English. Make assumptions explicit.”
Project-level example: “For this engagement, separate client requests from approved changes. Include a change log with revised proposal drafts.”
Keep general preferences broad enough to suit your wider work. Put engagement-specific procedures with the engagement. Give the immediate deliverable and any task-specific exception in the current request.
These placement suggestions are editorial guidance, not a claim about a universal instruction-priority hierarchy.
4. Memory: continuity that needs inspection
Memory can carry useful information between supported interactions. Past-chat search retrieves earlier discussions; it is a related but distinct capability.
Review and update remembered information when your circumstances or preferences change. For important commitments, consult the current source rather than relying on recall alone.
In our example: remembering that you are preparing a consulting proposal is helpful continuity. Establishing the approved fee requires the relevant approval evidence.
5. Projects: organize the engagement
Anthropic describes Projects as workspaces for related conversations, knowledge and instructions. Its current documentation also describes a new Projects beta rolling out in stages, so interfaces and behavior may differ between accounts.
For the fictional engagement, a practical workspace could contain:
- A short engagement brief.
- The original proposal, clearly marked as the baseline.
- Meeting notes and client requests.
- A dated approval record.
- Current drafts and an unresolved-questions list.
Organizing the files is only part of the work. Maintain clear status labels so a draft, a request and an approval remain distinguishable.
6. Project memory is not automatically isolated
Current documentation says Projects have their own memory, but account memory and project memory can interact. It provides a “Use account memory” setting for separating project memory.
Inspect the actual settings before assuming that a project name alone establishes an information boundary.
This is especially relevant when the same consultant works with several clients. A general preference may be reusable; a client’s confidential commercial terms need an appropriate scope and handling.
7. Cloud and local Cowork have different memory behavior
Anthropic currently documents shared memory between Chat and cloud-based Cowork. It says locally running Cowork sessions do not use that memory.
Consequently, “Claude remembered it in Chat” does not establish that the same information is available in every working environment. Supply essential task context where it is needed.
8. Turning off memory is not the same as deleting history
Current controls distinguish a chat with memory turned off from an incognito chat. The former can remain in conversation history; the latter is handled differently.
Also, deleting a conversation does not automatically remove memories derived from it. Review memory controls separately. Organization retention policies can affect these features.
Consult the current documentation before treating any setting as a complete deletion or confidentiality mechanism.
9. Keep a short engagement record
The following is a proposed working practice for our fictional case, rather than an official Claude feature.
Engagement: [name or identifier]
Current objective: [deliverable being prepared]
Approved decisions: [decision and source reference]
Superseded information: [what changed and why]
Open questions: [items awaiting a decision]
Latest reviewed output: [file and version]
Last verified: [date and reviewer]
Keep this record current and provide it when appropriate. It gives both you and Claude a concise reference that can be checked against the underlying documents.
10. A reusable continuation prompt
Continue work on [engagement/task]. Use the current engagement record and the documents I identify for this task. Before drafting: - Identify the current objective. - List the sources available to you. - Identify the approved decisions relevant to this task. - Flag conflicts, missing sources or potentially outdated information. Use remembered context as background. Verify material commitments against current source documents. If the evidence is unavailable, mark the point unresolved. Deliver: [required output] Review criteria: [observable checks]
Asking Claude to list its sources helps reveal gaps. Inspect that list and the resulting output; the prompt itself does not prove that retrieval was complete.
11. Your practical accomplishment
- Choose one ongoing, non-sensitive task.
- Write down what you expect Claude to carry forward.
- Classify each item as a preference, instruction, source fact or open decision.
- Inspect the relevant account and Project settings.
- Create a short current engagement record.
- Start a fresh conversation and check which sources and decisions the response actually uses.
Your accomplishment is an inspected continuity setup and a current task record. Record what the trial reveals about your particular account and environment.
Continue with Claude Complete
- Previous: #04 — How to Give Claude a Clear Task
- Start here: #01 — A Map of the Claude Ecosystem
- Continue to #06 topic: Claude Cowork’s Five Levels — previously published companion article
