Skip to content
Call wiki

Where should standing context live — memory, project instructions, or files?

Last updated 2026-08-17 · Confidence: documented — Anthropic’s memory docs and help center; the Cowork-instructions-are-CLAUDE.md equivalence is inferred.

Put standing context in instruction files you write and can read back — they load every session and never surprise you; treat automatic memory as a bonus, not a store.

Memory is a synthesis, not your past chats

Section titled “Memory is a synthesis, not your past chats”

Claude’s memory never replays old conversations. Current plans keep categorized entries Claude updates in real time; legacy enterprise memory is a synthesis of chat history refreshed every 24 hours. A new chat sees only those distilled entries — richer than a two-sentence summary, but still lossy and written without asking you. You can view, edit, pause, or reset it in settings, and incognito chats skip it entirely. Cowork projects add their own memory: context from earlier tasks can resurface in later ones, scoped to that project only.

Alejo recommends relying on explicit files instead — and often disables memory: a file you wrote is inspectable; a memory entry appears without telling you.

Instructions load every session

Section titled “Instructions load every session”

Project instructions apply to every chat or task in a project. Claude Code’s equivalent is CLAUDE.md: read in full at the start of every session — though delivered as a user message after the system prompt, so it steers rather than enforces. Static personal facts go one level up, in the user-scope ~/.claude/CLAUDE.md, which loads in every project. Whether Cowork’s Instructions field is literally a CLAUDE.md file on disk is undocumented.

The inclusion test is staleness, and the budget is small

Section titled “The inclusion test is staleness, and the budget is small”

The docs’ size guidance: target under 200 lines per file — everything loads regardless of length, but “shorter files produce better adherence.” A ~20k-token budget has no documented basis; the only 200-line/25KB cap in the docs applies to auto memory’s index file, not CLAUDE.md.

Alejo’s inclusion test: not “is this important?” but “will this go stale and confuse the model?” Sessions don’t see past chats, only the instructions — so bake decisions in as they’re made, in bullets, and prune finished-project state before an agent treats it as live. Bulk knowledge — downloaded context, navigation maps — belongs in plain files read on demand (see Drive search limits); repeatable procedures belong in Skills, which also load on demand.

Sources

Section titled “Sources”