Store decisions with sources and dates

An original research assistant could retain an agreed topic, the selected source URLs, and unanswered questions. It should distinguish a user preference from an inferred guess, and keep timestamps on information that may become outdated.

Preserve a path back to the underlying evidence

Summaries can omit an important qualification. Keep a way to recover the underlying source or task history before the agent relies on a remembered claim.

Decide what is worth keeping

A conversation already has a transcript. Memory is for facts that should survive after that transcript is trimmed: a preferred name, a project id, a constraint the user stated.

It is not a place to keep every tool result “just in case.” Unbounded memory becomes a second context window full of stale notes, and the model will treat a stale note as a fact.

Write memories as data, not as a diary the model edits in prose. A field with a source turn, a timestamp, and an owner can be corrected.

A paragraph that says “the user probably wants” cannot. If the user did not say it, do not store it as something they said.

Who can read and delete it

Memory that crosses chats can leak one project into another. Scope it to the user and the workspace.

Do not let one user’s agent read another user’s memory because they share a vector index and the filter was forgotten. Test that filter.

Deletion has to work. If a user removes a memory, later prompts must not still receive it through a cache, a summary, or a backup index.

Write that test. Also expire memories the product no longer needs.

A phone number kept forever is a privacy decision, whether or not you called it memory.

What to record before you rely on it

List the fields you store, the scope, and the expiry. Save a test where a deleted memory does not appear in the next prompt.

Do not store guesses. If a value came from a tool, store the tool result id so you can tell a user fact from a fetched page.

Keep the model id, the date, and the inputs you used beside that note, so a later change can be compared with the same test. When the result is a failure, store the artifact and change the design.

Do not relax the check until the model happens to pass. Name a person who owns that failure.

A note with no owner and no inputs cannot be repeated, and a pass with no date cannot be compared with the next model id.

Common questions

Is the chat transcript the memory?

It is the record of one conversation. Memory usually means something you inject into a later conversation.

Treat that injection as a feature with a schema, not as leftover text.

Should the agent decide what to remember?

It can propose a field. The product should constrain the schema and let the user see the result.

An unrestricted “remember this conversation” note will keep secrets and mistakes with equal care.

Sources & further reading

  1. Anthropic: Effective context engineering for AI agents ↗
How this story was made

Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.

Our editorial standards
Back to all stories