How agent memory should be captured

Fine-tuning your agent’s workspace setup can help agents know what to leave behind, so that upon refresh, Enzyme can make the next session better.

Good agent memory is not a private database the agent mutates after every chat. It is a trail of artifacts you can inspect: session notes, decisions, observations, preferences, research memos, handoff briefs, and project logs. The agent writes them down. You keep them in the workspace. Enzyme compiles them into working context for the next session.

A memory record that cannot be reviewed, moved, linked, corrected, or deleted outside the vendor dashboard is hard to trust. A markdown artifact can live beside the code, notes, customer research, or product corpus it refers to. It can be diffed. It can be cited. It can become part of the same structure humans already use.

Capture the trace, not just the conclusion

Most memory systems extract facts:

The user prefers concise answers.

That may be useful, but it loses the situation that made the preference true. Enzyme works best when agents capture the trace:

## 2026-06-01 session note

- User pushed back when the setup answer became a full content revamp.
- Preference: propose low-risk IA changes first, then implement for review.
- Boundary: do not overclaim fully local; catalyst generation may call the configured provider.
- Follow-up: update homepage around artifact-native memory, not generic memory layer.

The second version gives future agents material to reason with. It preserves the decision, the correction, the boundary, and the next action. Enzyme can index it, connect it to related notes, and surface it when the same tension reappears.

Prefer append logs over silent mutation

Append-based memory keeps time visible. A new project log file that says what changed this week is more useful than a single “current preferences” file rewritten until the history disappears. Agents should add new observations, not silently overwrite old ones.

Good writeback artifacts include:

• Session notes — what happened, what was decided, what remains open.
• Decision logs — options considered, chosen path, reason, date, owner.
• Observations — recurring user preferences or project constraints with evidence.
• Research memos — sources read, claims verified, open questions.
• Handoff briefs — current state, exact paths, validation, risks.

Give agents a writeback boundary

Do not ask an agent to “remember everything.” Tell it what it may write automatically and what requires review.

A practical boundary:

This keeps memory useful without turning every conversation into an unreviewed migration of your knowledge system.

Structure that helps Enzyme

Enzyme does not require a perfect vault. Incomplete structure is still signal. But a few conventions make the compiled context sharper:

• Date session notes and decision logs.
• Keep append-only logs in predictable folders such as inbox/, logs/, decisions/, or research/.
• Use wikilinks or stable names for recurring people, projects, and concepts.
• Tag sparingly where a tag expresses a real retrieval handle, not just a filing habit.
• Keep agent runtime folders (.agents/, .claude/, .pi/) separate from durable memory artifacts.

Enzyme doesn’t need beautiful organization. Rather, it benefits from enough durable shape for an agent to know where the trail lives and enough provenance for you to trust what it found.

Anti-patterns

Avoid structures that force every agent session to rediscover the whole world:

• One giant memory.md with unrelated facts and no dates.
• A people folder where every entity is treated the same regardless of activity.
• All notes in one folder with no temporal or project cues.
• Silent overwrites of “what we know” without preserving why it changed.
• Asking the model to organize everything from scratch each session.

Enzyme is strongest when memory is a source-controlled, reviewable artifact trail. The agent’s job is to write the trail, which Enzyme makes fast, associative, and useful next time.