Card 03 · AI agents, explained
AI agent memory: it remembers Friday on Monday
AI agent memory is what an agent keeps after a conversation ends: the facts, decisions and documents it can read next time without being told again. It is not RAM and not the model's context window; the test is whether Monday's answer knows what Friday settled.
01 — The handover
Friday's work meets Monday's blank page.
Ask a fresh chat what happened last week and it has nothing: the conversation ended and the context left with it. Memory is the difference between a chat that forgets and an assistant that picks up where the week stopped.
02 — What it keeps
It keeps what was decided, not every word.
Three facts beat four hundred messages: Acme moved to net-30, Leo covers when Priya is out, the Q3 timesheet is still missing. That is what you would write on an index card, and it is enough to start the week.
03 — Where it lives
It lives in a file you can open.
Good memory is plain text you can read and correct, not a number buried in a model. When a fact is wrong you edit the sentence, and every later run reads the correction.
04 — The date
Every fact carries the day it was true.
Acme was net-30 in October, and in January it may not be. A dated fact lets Monday's run show its work, and lets a person spot the stale one before the client does.
05 — Without it
Without it, the morning goes on re-reading.
The agent opens the old threads and rebuilds the week from half-answers and screenshots. A team of ten spends the morning doing Friday's work a second time.
06 — The open question
What it does not know is part of it.
The missing timesheet is written down as missing. That line is why Friday's report cannot claim a number nobody has, which is the failure people actually get burned by.
07 — The owner
Personal memory belongs to one person.
In TalentOS a personal chat, its memory and its approvals are owner-only, so what Priya settles with her agent stays hers. Work the team needs lives in the Room, where everyone's agent can read it.
08 — Start here
Write down what Friday settled, one line each.
Take the three facts your team repeats every week and put them where every agent reads them. The gain is not the remembering; it is that Monday starts at the second step.
Read on Monday morning, before the report is written.
It bills Acme net-30, flags the missing timesheet and asks for it before any number goes out.
Tools for this
Every price and licence below is what the tool’s own page said on 18 November 2026, with two sources behind each record. There are no scores and no order of preference; where TalentOS is one of the tools it is marked ours and listed like the rest.
A memory layer that keeps what an assistant learns as plain Markdown notes in a local folder, read back through MCP by any compatible assistant.
For a small team that wants its AI memory to be files it owns, can open and can move.
$0 self-hosted, cloud from $19 per month · Open source, AGPL-3.0
Basic Memory repository and READMEBasic Memory plans and prices
A workspace where one assistant answers the whole team in a shared thread, backed by a memory of Markdown files that every connected AI can read.
For a team of about 5 to 15 people that wants one shared assistant plus one memory, and wants that memory kept as plain files it owns.
See its pricing page · Open source, MIT
A platform for stateful agents: agents carry memory blocks you can read and edit, and run locally or on Letta's cloud.
For a developer who wants agents that remember between sessions without writing their own memory store.
$20 per month (seat) · Open source, Apache-2.0
A memory layer for AI apps that pulls facts out of a conversation, stores them, and returns the relevant ones on a later request.
For a developer adding long-term memory to an agent or app without building a store and a retrieval pipeline.
$0 with your own model key (free) · Open source, Apache-2.0
A hosted memory and context engine that stores what an assistant should remember and puts the relevant part back into a later request.
For a developer who wants a managed memory API and would rather not run a store and a retrieval model.
$19 per month (usage) · Open source, MIT
Supermemory plans and creditsSupermemory repository and licence
A context and memory service for agents built on a knowledge graph that tracks when a fact was true, sold as Zep Cloud with an open-source graph library underneath.
For a team whose agents need to know what was true when, not just which text looks similar.
$125 per month (usage) · Open source, Apache-2.0