CLAWPOD ENGINEERING
Continue Long Work with Verifiable Memory and Handoff Notes
Separate conversation context, durable memory, and work status. Use a filled handoff example to verify decisions and recover when records are missing.
At the end of a long conversation, “do what we agreed earlier” makes it hard to check which decisions and sources the Agent is using. This guide is for people who assign research and reporting work across several days in ClawPod Cloud. You will learn to separate the current conversation context, durable memory, and unfinished work records, then verify the starting point for the next step. You need access to the Agent and source material; memory and work-record requests remain subject to available tools, write permissions, and organizational policy.
The goal is a work record you can verify, not simply a conversation you can keep extending. Record important decisions with their evidence, then confirm that the Agent actually read those records before continuing. A request to remember something, or a single summary, does not guarantee complete retention or automatic resumption.
Conversation history, current context, and durable memory differ
Context is the information a model uses together to produce its current response. A session is a unit in which the runtime manages conversation and execution state; do not treat it as identical to a Room on screen. Memory records lasting decisions, preferences, and work context in a managed Agent workspace so they can be retrieved when needed.
The official Guide says that long bootstrap documents—the instructions supplied at startup—can be truncated by context limits. That does not mean conversation messages can never be shortened or summarized. Nor does seeing an old message on screen establish that the model is using the entire message in its current response.
The Guide pages reviewed for this article do not establish an exact context size for every Agent, a trigger for automatic summarization within the same session, or a guarantee that every detail survives a summary. Do not design work around assumptions such as “there are one million tokens, so we can keep adding material” or “automatic summaries make handoff notes unnecessary.”
Summarizing and saving require different checks
Summarizing older conversation turns to reduce the current context is commonly called compaction. Here, the term distinguishes a concept; it is not an instruction for a particular ClawPod Cloud setting or screen. Even if an installed Agent runtime provides this behavior, that alone does not verify your organization’s configuration, a successful operation, or a retention period.
What the Guide explicitly documents is that recent context can be saved as daily memory when a new session starts or a reset occurs. Do not equate that possibility with automatic compaction inside the same session, complete transcript retention, or unconditional resumption of work. Whether a daily record was created and whether the next response read the relevant part are separate questions.
For important work, first ask the Agent to report which records it read, which decisions remain current, and what it could not verify. A connection to the original material is a better check than a fluent summary.
Put memory and work status in different records
The following applies the Guide’s role distinctions to everyday work. These records do not automatically replicate or replace one another.
| Information to retain | Suitable record | What to check next time |
|---|---|---|
| Lasting preferences and confirmed decisions | Memory | Current wording and change date |
| Evidence and detailed research notes for this job | An approved detailed document | Location, revision, and source links |
| Team ownership, progress, and acceptance criteria | Tasks | Work status and linked deliverables |
| Execution steps and status managed by the Agent | Workboard | Card results and unfinished work |
Tasks tracks shared work across an organization. Workboard manages cards and execution state on the Agent’s work board. In either case, marking an item complete does not establish the accuracy of its deliverable. Choosing a work board is also different from requesting parallel execution.
Memory search depends on the runtime’s memory tools and workspace policy. The Guide says to search first and read only the relevant portion; saving also requires the applicable tool policy and write permissions. Separate an acknowledgment of “remember this” from reading back the saved content.
A request for handing off a multi-day report
The following is an illustrative example, not customer data or an observed run. It assumes that the required tools and source access are permitted. The aim is to draft a comparison of three internal operations reports without modifying the originals or sending them outside the organization.
Compare the April, May, and June operations reports in the approved working folder and create a draft for internal review. Do not modify the originals or send them externally. Compare processed volume and average processing time; exclude cost comparisons.
First confirm the criteria and source revisions. Record acceptance criteria, deliverable locations, and open questions in the work record. When stopping today, write a handoff note listing the sources read, verified results, unknowns, and next action.
Retain only this team convention in durable memory: “Monthly operations comparisons check processed volume and average processing time first.” Do not copy report originals or temporary calculations into ordinary memory. If an existing convention conflicts, identify the conflicting record and do not overwrite it before clarification. If memory or work-record tools are unavailable, do not claim completion; report the deliverable you can provide and the limitation.
This request separates a lasting convention from the details of one execution. It specifies what to save, but it does not itself establish that saving succeeded or that a future response will retrieve it automatically.
What a useful handoff result looks like
Illustrative result—not execution evidence. Ask for a filled record like this and compare it with the original sources. The filenames are examples too.
| Item | Sample result |
|---|---|
| Sources | Read revision 2 of the April and May reports and revision 1 of the June report |
| Verified finding | Processed volume has the same definition in all three. June’s average processing time needs a footnote check |
| Deliverables | operations-comparison-draft.md and handoff.md in the approved working folder |
| Open question | Does June’s average include waiting time? |
| Work status | Draft written; metric comparability remains incomplete. Not marked finally complete |
| Memory check | Saved the single team convention, then read it back and confirmed a match. Did not save report originals |
| Next action | Ask the responsible person about June’s definition, then determine whether a like-for-like comparison is possible |
The important result is not a convincing final report. It is keeping the unresolved comparison visibly incomplete. If an actual response omits the record location or verification evidence, do not assume it reached the final checks shown in this sample.
Check four things before the next step
- Ask the Agent to read the handoff note and the specified original revisions, then report the evidence locations. A claim to remember the earlier conversation is not a substitute.
- Confirm that cost remains excluded and the restrictions on original-file changes and external transmission still apply.
- If June’s definition remains unknown, do not accept a definitive numerical comparison. If an answer is available, record its source and confirmation date and distinguish it from the previous assumption.
- Compare the deliverable with the acceptance criteria before updating work status. Separately read the durable memory to confirm that it contains only the team convention.
These checks concern the specified records and deliverables. Do not expand them into a privacy guarantee that particular content has been completely removed from every store, log, and backup.
Recover from missing or conflicting memory using verifiable sources
Do not invent an earlier decision because a search returned nothing. Check whether memory search is available and whether the relevant records can be read, then return to the approved handoff note and original material. If a document is inaccessible, have the Agent identify the missing source and required access, and continue only work that does not depend on it.
If an earlier record says “compare cost too” but the current request excludes cost, surface the conflict. Confirm which instruction is current and record the change date and evidence. Do not infer permissions or approvals from memory either.
When a tool failure leaves saving uncertain, do not report success. Where possible, inspect the actual state through a safe read before taking further action. Do not blindly repeat changes that could create duplicate records or unintended deletion. For deletion, confirm the target and scope first and inspect the actual result. Writing a date in a retention request does not itself schedule automatic deletion.
Short instructions and explicit limits provide a better starting point
The Guide recommends one responsibility per workspace file and short, verifiable instructions. Separate role and tone, recurring work rules, and tool notes; link longer supporting material from separate documents. Putting important rules first is an authoring practice, not a guarantee that a particular sentence will be injected or every restriction automatically enforced.
Keep these boundaries as well:
- Do not put passwords, tokens, API keys, or unnecessary customer originals in ordinary Memory or work documents. Handle secrets through a dedicated path approved by your organization.
- A managed workspace does not by itself establish retention periods, complete deletion, or a level of privacy protection. Check your organization’s policies.
- A feature in runtime documentation does not establish a particular menu in the Cloud Portal or Agent Control UI. This article is not a settings walkthrough verified against the current UI.
- Memory, summaries, and work records do not replace tool permissions or approval for external changes.
For your next long-running job, request a handoff note with the goal, current decisions, evidence locations, unfinished items, and next action. The handoff is not finished until the next step reads that note and checks it against the originals.
Sources
- ClawPod Guide — Memory and Agent settings: memory-tool conditions, long startup instructions, and possible daily records at a new session or reset.
- ClawPod Guide — MEMORY.md and the memory folder: lasting and detailed records, read/write verification, and content not to store.
- ClawPod Guide — Agent workspace files: file responsibilities, concise verifiable instructions, and boundaries around automatic execution and permissions.
- ClawPod Guide — Tasks and Milestones: shared work status and checking deliverables before completion.
- ClawPod Guide — Chat modes: Tasks and Workboard roles, plus execution, review, and permission limits.
