Context engineering: sources, memory and handoffs for resuming work

For this blog, I keep sources, research and articles in Notion. A source can support several pieces and become useful again months later, when its findings still hold but the tools or conditions have changed. Saving the link makes it findable. Reusing it also requires keeping what it says, the setting it describes and where I have used it.

METR’s publications on developer productivity make a useful case. The 2025 finding, the 2026 methodological updates and the survey of perceived benefits answer different questions. A hasty summary can flatten them into “AI slows developers down”. That sentence then enters another article and loses its connection to the research.

This makes context engineering a concrete problem: choosing which information an agent should receive for its next step, including instructions, documents, tools and work history. Context is what the model receives at that moment; a page stored in a wiki has to be retrieved to become part of it. Anthropic, Effective context engineering for AI agents.

Keep the question alongside the source

In July 2025, METR published a randomised study involving sixteen experienced developers and 246 tasks from their own open-source projects. With early-2025 AI tools, they took 19% longer. The finding concerns that sample and those conditions. Storing it as a forecast for my current tools would change the meaning of the study. METR’s July 2025 research.

In the Resources AI register, the source record keeps the finding alongside its scope. This is a reduced transcription of the fields present in the record, with their contents rendered in English:

The field I care about most is the reuse limit. A number is easy to remember; its conditions tend to disappear as a note passes between summaries. Keeping them together makes it harder to transfer the 19% figure to a setting the study did not measure.

Reading coverage helps determine when to return to the original. Reading the results and discussion differs from examining every detail of the paper. The date records when the material was checked; it does not make the tools studied current.

To reproduce this organisation in Notion, create a Sources collection with one record per publication. URL, topic and checking date become searchable, filterable properties; the question, passage read and reuse limit stay in the record’s body, with room for explanation. Before adding a source, I search for its URL to avoid maintaining two syntheses of the same document.

Then create a page for the research question and link it to the source records. In my case, that question concerns adopting agents in a team. The page holds the useful conclusion and the reasoning supporting it. A link to the article identifies where the conclusion is reused. The collection and these linked pages are enough to start; I would add properties when a search I need becomes difficult.

The distinction matters when something changes: the source record captures what I read, the synthesis holds what I infer, and the task records what remains to do. A generic “updated” field would not tell me which of these happened.

From finding to decision, with the reasoning intact

In the research for the team-agents article revision, the METR record is linked to the dossier and both candidate texts. The useful synthesis is that feeling faster at producing code can diverge from the time needed to complete the work. To decide on adoption, I also want to observe verification, corrections and human intervention.

The register also contains two later sources. In February 2026, METR describes problems with its new experiment: the people willing to participate and the tasks selected change, while parallel-agent work complicates time accounting. The authors consider those data an unreliable signal of the current effect. The survey published in May instead collects perceived benefits, with the limits of a convenience sample and self-reported estimates.

The 2025 finding remains unchanged. What narrows is the inference I could draw today: I cannot choose a current workflow by applying that slowdown to it, or replace it with the survey’s perceived speedup.

In Notion, all three sources appear in the productivity section of the Team article. This is the path I want to reconstruct: original publication, source record with setting and limits, research synthesis, reuse in the text.

When new evidence arrives, I start with the question it might change. If it concerns one client’s implementation, I check the passages about that client. If it changes how time is measured, I return to the syntheses and articles using those measurements. A new source alone is no reason to rewrite every linked page.

What goes into the session

To revise that productivity paragraph, I retrieve the article’s question, the passage to change and the three METR records. I then open the original sources where the comparison requires it. The history of every editorial discussion would add material without clarifying the claim I am checking.

To make that retrieval actionable, I would use an instruction like this. The public references allow the sources to be opened; the current article paragraph must be attached or read through its link.

Task: review the productivity paragraph in the team-adoption article.

Read the current paragraph and these sources:
- METR, July 2025 experiment:
  https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
- METR, February 2026 methodology update:
  https://metr.org/blog/2026-02-24-uplift-update/
- METR, May 2026 survey:
  https://metr.org/blog/2026-05-11-ai-usage-survey/

Question: does the paragraph distinguish the historical measured
result from evidence about current tools?

Compare the claims, their populations and what was measured.
Return a proposed edit only where a source changes a claim.
If the paragraph is already accurate, explain why it should stay.
Keep unresolved questions visible.

The final question narrows the reading: I am looking for an unjustified transfer of a finding, rather than a summary of everything METR has published. If the comparison needs a detail missing from the record, I open that part of the original. The note helps choose where to read.

This is the boundary between active context and retrievable knowledge. The former contains what the current decision needs; the latter remains reachable through precise references. Anthropic describes retrieval as needed as one strategy for managing context. Context retrieval and agentic search.

A useful reference suggests what the reader will find. “METR, task selection and time with parallel agents, February 2026” guides a search better than “interesting link”.

Retrieval can also take the wrong path. Keeping the question and its supporting passage in the synthesis helps expose a document that is relevant to a different question.

A wiki preserves knowledge; a handoff preserves a work checkpoint

A source can help several articles. The next step in a revision belongs to a specific task. Mixing these layers creates pages full of expired instructions.

History preserves how a choice developed, including abandoned attempts. A wiki collects reusable knowledge and its sources. A handoff describes where to continue interrupted work.

To pause the METR paragraph comparison, I would use this handoff. It is an example built around the same material, to be completed with the link and version of the working text.

Objective: check the METR paragraph in the team-adoption article.
Current material: link to the working text, revision and date.
Source records: July 2025 experiment, February 2026 update,
May 2026 survey, linked from the research page.

Established: the 19% slowdown belongs to the 2025 experiment.
Open: does the current paragraph extend that finding to today's tools?
Next action: open the current text and compare that claim
with the three source records; read the originals where needed.

Recheck if: the paragraph, the question or a relevant source changes.
Expected result: a justified keep/edit decision, with sources.

“Current material” must point to the actual text and the revision read. A modification date can reveal a change; understanding it requires comparing versions. For code, the equivalent reference is the commit together with any uncommitted diff: a test result needs to stay linked to the code checked.

On resumption, I would follow this order:

1. Open the current text. Check whether the paragraph matches the one left at the pause and whether the request has changed. If I now need to assess 2026 tools, the 2025 study record alone no longer covers the question.
2. Reread the research conclusion and the relevant sources. The 19% remains a historical finding. The February update concerns difficulties with the new experiment; the May survey measures perceptions. A newer source does not automatically replace an earlier one.
3. Decide on the specific passage. If it attributes the slowdown to that sample, keep it. If it extends the finding to current agents, correct that extension. When a relevant measurement is missing, leave the estimate of current benefit open.
4. Update the work checkpoint. Record the passage checked, sources used, decision and remaining questions. If the synthesis changes, follow the links to other articles reusing it and inspect the affected passages.

The Team paragraph linked above already distinguishes the 2025 finding from later sources. Reopening the sources can confirm a choice: that passage can stay. How useful an agent is today remains a separate question, requiring observations of the work I want to assess. The full comparison is in the team-adoption article.

In Anthropic’s long-running agent experiment, a new session reads progress and history and checks the application’s operation before continuing. That is the step I would retain in another environment: comparing the checkpoint with what exists now. An old “test passed” remains a record of that check; a changed contract may call for a different test.

Compaction and a new session serve different purposes

Compaction summarises a conversation so work can continue with less active context. Persistent notes remain outside the model’s window and are retrieved when needed. Aggressive summarisation can lose the detail on which the next step depends. Compaction and structured note-taking.

For METR, I would shorten the discussion and repetitive outputs while preserving the question, sources and distinction between a historical result, a current measurement and a perceived benefit. The original passage stays reachable through the register. The open comparison belongs in the handoff, rather than being copied into every wiki record.

Delegation also separates contexts. I can ask an agent to examine a study’s conditions and return its conclusion, relevant passages and open questions. That report enters the main context, while the detailed exploration stays in the delegated session. To use a conclusion, I need a path back to its evidence: removing detail from the input does not mean deleting it from the archive.

Opening a new session makes retrieval explicit. I would choose it when the task changes and the old conversation mostly contains superseded attempts. Staying in the same conversation may be more practical while a line of reasoning still depends on points just discussed. The choice needs testing over the whole task: writing the handoff and rereading sources also take time and resources.

Either way, restarting with a different context leaves existing external effects in place. Before resuming a change, I check the repository or service involved rather than infer its state from the conversation.

An LLM wiki needs maintenance

In his LLM Wiki proposal, Andrej Karpathy describes a knowledge base in which an agent integrates new sources into linked pages and maintained syntheses. Sources, derived knowledge and maintenance rules have distinct roles. In the editorial system in Notion, the register keeps sources, research develops the reasoning, and articles use its relevant conclusions.

Maintenance matters as much as the initial synthesis. A duplicate page can preserve an old conclusion while the updated one lives elsewhere. Before creating another record, I search for the topic and URL; when a conclusion changes, I also check the uses it could affect. A commented link can suffice for a source used only once.

A wiki can coexist with retrieval-augmented generation: generation using information retrieved from an external knowledge base. Lewis and colleagues’ RAG work combines generation with retrievable memory. A synthesis page can itself be retrieved, then point to the original document for detail. RAG describes retrieval and generation; the wiki adds a maintained organisation of knowledge.

For the METR question, I would start from the synthesis, check the limit that changes the decision and open the relevant source. For a new question absent from that synthesis, I would return to research. Retrieved pages remain material to evaluate: text inside them grants no permission to send data or change services.

If the handoff sends me through dozens of pages before identifying the open passage, I narrow it. The knowledge base can be broad; the point from which I resume needs to be recognisable.

Preparing and maintaining these notes is part of the work’s cost. I also account for that in my Claude and Codex cost research: a new session offers little if making it usable means reconstructing everything. Here, I judge the structure by the work those notes let me continue.