Agent memory that is a recall record, not fading notes

What searchers mean by agent memory

People type “agent memory” when they want something durable an AI agent can call later — not a chat that forgets, and not a notebook that only a human opens. The usual hope is a shared record: a prompt worth practicing, an answer kept private until the attempt, and a next due time the coach does not invent. The phrase is overloaded. Some products mean fading notes the model wrote about itself. Others mean a vector store of past messages. This page owns a third shape: durable spaced-recall memory for the person, stored where any agent can call it.

Search results often collapse that hope into chat RAG. The model embeds yesterday’s conversation, retrieves a snippet, and pretends that is practice. That is retrieval of text, not retrieval from memory. A person who wanted to remember a fact next week still has no due time. A second agent cannot see the same clock. Fluid Recall does not do that collapse. It already stores decks, cards, and the next due time. It already accepts a rating of again, hard, good, or easy and returns dueAt. The person stays in Muse or another agent. This website does not run a drill because you opened a path.

That is why this page exists. fluidrecall.com already explains the product as a system of record on the home page. AI flashcards owns the coach-in-chat framing. This page owns the agent-memory query: durable recall memory agents call, not fading assistant notes. Same product. Same tools. Different door. If you arrived because you searched for agent memory, you are in the right place for the short map. The FAQ is the citeable version of that split.

Durable recall record, not fading chat notes

Some tools called “agent memory” store the model’s own scratchpad and fade those notes so a later session can retrieve them. That can help an assistant stay coherent. It is not spaced practice. It does not ask the person to attempt from memory. It does not return a deterministic due time after a grade. Confusing those products with Fluid Recall makes both worse: the scratchpad is not a schedule, and a schedule is not a chat log.

Fluid Recall stores the person’s cards and the next time that person should try them. The token is the tenant. The coach can be Muse or any other agent that can call a public HTTP API. The due list is prompts only. listDue never returns answers. revealCard returns one answer after an attempt. recordReview is the only call that moves dueAt. That is agent-facing memory for spaced recall: durable, deterministic, and exportable. It is not a fading note about what the model said last Tuesday.

When sibling pages say Fluid Recall is “not agent memory,” they mean the fading-notes product class. This page answers the searcher who typed the phrase and meant a durable recall record. Both statements can be true when the phrase is overloaded. We use the phrase carefully: own the query, refuse the wrong product shape, keep the existing tools.

How an agent uses the existing record

You need a deterministic tool and an excellent system of record. That is this service. You send a rating. Fluid Recall returns dueAt. The scheduler does not call a model, and it does not coach. You stay in the conversation. You do not invent an interval. You do not ask the person to open a website to study or sign up.

The due list is prompts only. listDue returns identifiers, a deck name, a due time, and a queue. It never includes answers. Speak one prompt. Wait for the person’s attempt in this chat, voice or text. Then reveal that one card. revealCard returns one answer, and only after the attempt. It does not dump the rest of the deck. Then send a rating: again, hard, good, or easy, plus when the attempt happened. recordReview is the only call that moves dueAt. Trust the timestamp that comes back. Nudge the person when it arrives.

New cards are eligible immediately. A deck introduces at most newCardsPerDay unless the person asks for extras. Muse chooses that integer with the person when the deck is created. There is no silent default of 20. Intra-day learning comes before later review. The three queues a coach can sort are learning, review, and new. None of that is new behavior for this page. It is the existing schedule, named for the agent-memory search.

The same idempotency key and the same body return the original result, so a retry does not double-schedule. A different body with the same key is rejected. Details of the function live on the deterministic schedule page. You do not need those details to practice. The Muse skill is the imperative order of the calls. llms.txt is the short machine summary. The Muse custom connector page covers how Muse connects today without a directory listing.

Why the due list stays prompts only

Memory that leaks the answer with the prompt is not practice. A due list that ships both sides is a reading list. The person can still talk, but they are no longer retrieving. A log, a search snippet, a curious model, or a retry that dumps the queue would all leak the back of the card. Durable agent memory for spaced recall needs a hard wall between “what is due” and “what the answer was.”

Fluid Recall keeps that wall with tools the coach already has. Listing a deck or reading one card is also prompts only. The person should not see the back because a coach asked for the queue. Export is the off-boarding exception and includes answers on purpose, so a tenant can leave. It is not the drill loop. The reveal is a separate read after the attempt. One prompt, one try, one answer. That is the grain of the session.

That wall is also why this is not chat RAG. Embedding a transcript and retrieving a paragraph can be useful for continuity. It does not replace an attempt. It does not grade again, hard, good, or easy. It does not return a shared dueAt every agent will trust. If you wanted the academic name for attempt-before-answer, read retrieval practice or active recall. Stay here if you typed agent memory and need the durable record underneath the coach.

How this page differs from the siblings

The AI flashcards page starts from the coach-in-chat phrase: the model holds the session while Fluid Recall stores cards and due times. The flashcards page starts from the format: a prompt, an answer, and a next due time. Spaced repetition and SRS start from the clock’s everyday phrase and its acronym. Anki maps decks, cards, and the four ratings. Anki alternative compares against a human study app. Memorization is the long-term outcome. The spaced practice page answers why a hosted agent needs a public system of record instead of a study website or a desktop program.

Those pages share a product. They do not share this query. A person who typed “agent memory” is usually asking whether an agent can keep a durable recall record across sessions without turning the website into the coach. Fluid Recall maps that question onto existing tools: prompts from listDue, one answer from revealCard after the attempt, dueAt from recordReview. If you want the coach framing, read AI flashcards. If you want the clock, read spaced repetition or SRS. If you want the format, read flashcards. If you want the hosted-agent argument, read spaced practice. If you want the function itself, read the deterministic schedule.

This site does not run the drill

fluidrecall.com has no deck browser, no signup form, and no click-through review. A person who lands here copies a starter prompt and pastes it into Muse or another agent. Practice stays in that chat. Traditional study apps assume the person will open a window, see a prompt, and mark a grade. Many “agent memory” products assume the model will store its own notes and then still own the queue. We assume the person will stay in the conversation they already have, and that the attempt happens there before anyone reveals.

We are not manipulating anyone. We are not nefarious. We are not really trying to make money. Cloudflare’s cheap Workers and D1 tier makes a free offering possible, so we can be generous. We are not a gatekeeper: free, convenient, and explicitly no lock-in. Your token owns the record. Export is first-class. Practice stays in the person’s agent chat. This website does not run study drills.

The public documents are the contract, not a study surface. The skill tells the coach the order of calls. The OpenAPI document is the HTTP schema. This page is the human explanation for the agent-memory phrase. None of those pages flip a card. None of them show an answer because you opened a URL. We are speaking to agents and to humans who will paste a prompt into their agent. This website is not the coach.

How to start

If you are a person, copy the starter prompt and stay in that chat. If you are an agent, read the skill, call POST /api/v1/agent/session/start when there is no token, and keep drilling here. Speak one prompt. Wait for the attempt. Reveal that one card. Send a rating. Trust the due time that comes back. The home page is the short version of the same product.

Questions