Tokens, Context, and Memory: What an Agent Can Use

Tokens are units used to represent model input and output; they are not a dependable word count. The context window holds the material available for a request, including tool activity and outputs. Summarization can compress earlier work but does not guarantee every detail survives. This is distinct from model training.

Have a short fictional change log and the current version of the contact form. Use invented email addresses such as [email protected]. Do not use a full customer conversation as practice material.

A long conversation began with “do not change the delivery endpoint.” After many screenshots and test logs, the next request says only “finish the form.” Instead of assuming that the old boundary remains salient, supply a compact restart packet.

• Current goal: persistent email label; endpoint unchanged.
• Current state: label added in contact.html; keyboard test still pending.
• Evidence: local diff reviewed; no delivery request sent.
• Next action: verify tab order and 390px layout; ask before any endpoint change.

1. Keep stable project facts in an intentionally maintained instruction file, not a sprawling transcript.
2. Put task-specific decisions, unfinished checks, and actual file paths in a handoff note.
3. Link to a large test log and extract only the relevant failure, command, and environment. Do not paste thousands of unrelated passing lines.
4. Ask the next agent to restate the goal and boundary, then verify those against the files before editing.

A good restart packet makes the next step unambiguous while preserving uncertainty. “Browser check not run” is more useful than a summary that compresses pending work into “tested.” Claude Code’s authored CLAUDE.md and its auto-memory mechanism are distinct; neither should be treated as a permission barrier.

If the agent refers to a removed file or obsolete plan, stop the edit and reconcile the state. If a summary omits a safety boundary, repair the packet before continuing. More pasted history is not automatically better context.

• Does each state claim match the current checkout?
• Are pending checks visibly different from passed checks?
• Could a fresh reader find the source artifact without guessing what “that file” means?