Why agents forget between sessions, and what to save

A language model has no memory of its own between requests. Everything it knows about your project in a given moment is whatever text was placed in front of it: the instructions, the conversation so far, and any documents your code added. When the session ends, that text is gone. The next session begins with the model's general training and nothing else.

A one-off question survives that. For an agent working on the same project for weeks, it is the main source of friction. Rules get pasted in again. Decisions get made twice, sometimes in opposite directions. A review agent cannot see why a trading agent rejected the same setup on Tuesday.

The context window is not a memory

Long context windows help within a session, but they are a desk, not a filing cabinet. They are cleared when the session ends, they are expensive to fill, and models pay less attention to the middle of a very long prompt. Memory is the part you choose to keep and bring back on purpose.

What is worth saving

• Facts that stay true: names, addresses, the resolution rule of a market, the contract a token lives at.
• Preferences: tone, format, units, limits a person set.
• Decisions with reasons: what was chosen, what was rejected and why. The reason is the part that saves time later.
• Sources: a link and a date, so a later session can tell old evidence from new.

What to leave out

Raw transcripts are tempting but rarely useful: they are long, repetitive and hard to search. Secrets do not belong in plain-text memory at all. If something private has to be remembered, encrypt it in the browser first, which Pinlo does with a passphrase that never leaves your device.

Small records recall better

Recall in Pinlo matches words and tags. It is deliberately simple, so you can predict what comes back. That works best when each record says one thing. "Prices are quoted in USDG" is easy to find; a page of notes that mentions prices once is not.

To try it, open a mind in the app, save three records, reload the page and recall them. Then give a second agent a read grant and watch it pick up the same thread.