Lesson: How Clawless remembers (and forgets)

Imagine restarting every conversation by re-explaining who you are, what you do, what you are currently working on, and what kind of answers you prefer. Every time. Every single time.

That is what AI without memory looks like. You sit down, you type, the agent helps with the immediate question, and the next time you open the app it has no idea who you are. The same exchange you had yesterday plays out again from scratch. Useful, but exhausting.

Clawless remembers. The chapter you are about to read is the one that explains how, and how to stay in control of it.

About eleven minutes to read.

What memory actually is in Clawless

The first thing worth being careful about is the difference between two things that sound similar.

Conversation history is the literal back-and-forth of a single chat. Every message you sent in this conversation, every reply the agent gave, in order, from the top. Conversation history is bulky and lives inside one specific conversation. It does not travel anywhere on its own.

Memory is the distilled facts pulled out of those conversations over time. Your name. Your job. Your current projects. Your preferences about how answers should be written. Memory is compact, and it travels with you across conversations and agents.

History is what happened. Memory is what is worth remembering.

The Clawless engine has a quiet classifier that runs as you talk, watching each turn for things that look worth keeping. Names. Preferences. Decisions. Ongoing projects. Important dates. If it catches one, it pulls the fact out of the conversation, writes it as a short standalone memory, and sorts it into a tier based on how important it looks.

You will not see this happening most of the time. The classifier runs in the background. You will see the results, in the Memory panel, the next time you look.

The four tiers

Clawless sorts memories into four tiers, from most-important to least. The tiers are not a ranking system you have to manage. They are how the engine decides which memories to include in which conversations.

Pinned

Pinned memories are the facts you have explicitly told Clawless to always remember. These follow you across every conversation, every agent, every day. They are the most load-bearing tier.

Typical examples look like:

My name is Sam.
I run a small consulting firm focused on supply-chain analytics.
I prefer concise answers under 200 words unless I ask for more.

These are the kinds of facts you do not want to retype every time you sit down. The tradeoff is that Pinned content goes into every prompt the agent sees, which costs a few extra tokens per message. So Pinned has a hard size budget (around 2,200 characters total). When you try to pin more than fits, Clawless asks you to consolidate or remove a few existing pins before adding the new one. The cap is intentional. It is what keeps the “always remember” tier fast and cheap.

Insights

Insights are the things the classifier inferred from your conversations and is confident enough about to surface routinely. Not as load-bearing as Pinned, but useful enough to feed back regularly.

You do not have to do anything to create Insights. They build up on their own. When you look at the Memory panel, you can review which Insights the engine has gathered and promote the important ones up to Pinned with a click.

General

Everything else the classifier has noticed but is not sure is important. General memories are included in conversations only when they look relevant to what you are currently asking about. They do not weigh down every message; they show up when they fit.

Decayed

Decayed is the path to forgetting without deleting.

If a memory has not been touched in a long time, it ages into Decayed automatically. It still exists, you can still see it in the Memory panel, but Clawless does not surface it in new conversations unless something makes it look very relevant again. If you stop working on the project a memory was about, the memory fades. If you come back to that project a year later and start mentioning it again, the memory wakes back up.

This is the difference between human forgetting and computer deleting. Humans do not delete memories; we simply stop accessing them. Decayed mimics that. The fact is not gone. It is just resting.

How memories get in

Three pathways.

The classifier picks them up automatically. This is the default. As you talk, the engine watches for things worth keeping and pulls them out into memory candidates. Most of what ends up in Insights and General arrives this way. The quality depends on the AI provider you have selected, because the classification runs server-side on whichever model you are using.

You ask explicitly. You can tell any agent something like “remember that I prefer answers in three bullet points” and the agent surfaces it as a memory candidate. You confirm before it is saved. This is the right path when you have a specific preference the classifier has not picked up on its own.

You add one manually. In the Memory panel, click Add Memory and type the fact straight in. Useful when you want to seed memory at the start of using Clawless (“My name is Sam, I run a small consulting firm”) without waiting for the classifier to figure it out.

Once a memory is in the system, regardless of which path it took to get there, you can edit it, pin it, unpin it, move it between tiers, or delete it.

The Memory panel

Open Memory from the navigation rail on the far left. The Memory panel opens, showing every memory Clawless has saved, grouped by tier, with the most-important tier (Pinned) at the top.

For each memory you can:

Pin or unpin to move it into or out of the Pinned tier.
Edit to rewrite the text. Useful when the classifier saved something slightly wrong and you want to clean it up.
Delete to remove the memory permanently.
See where it came from. Each memory carries a small badge showing which conversation it was extracted from.

A search box at the top filters across all tiers by keyword. Type “name” and you see every memory that contains the word name. Type a project’s name and you see what the engine has remembered about that project.

The last-seen date on each entry tells you the last time the memory was used or referenced. Old dates are how you spot candidates for the Decayed tier; new dates are how you confirm an Insight is actually being relied on.

By default, all of your agents share the same memory pool. If you tell the Assistant your name, the Writer knows it too. The Researcher knows it. The Tutor knows it. There is one pool, and every agent draws from it.

This is the right default for most people. The whole point of memory is to stop re-explaining yourself, and re-explaining yourself per-agent is almost as exhausting as re-explaining yourself per-conversation.

When you switch agents in the middle of a task, the new agent does receive a short briefing of what was happening in the other agent’s conversation, automatically. Think of it as the difference between switching apps and switching colleagues. The new colleague has not been in the room, but you give them a one-line summary so they can pick up the thread.

Per-agent memory isolation, where one agent forgets while another remembers, is on the roadmap. At launch, all agents on a single Clawless install share the same memory pool.

Privacy: where memories live

Two facts to internalize.

Memories live on your computer. The Memory panel reads from a file in your local Clawless data folder. The memories are not uploaded to any Clawless server. If you turn off your computer, your memories sit there until you turn it back on. If Clawless went out of business tomorrow, your memories would still be there, in the same place, readable by you.

Memories travel with messages to the provider. When you send a message to an AI provider, the relevant memories travel along with it as part of the prompt. This is how the agent knows your name on a fresh conversation: your name comes along with every message you send. So the provider does see your memories, at least the ones that travel with each specific message.

The practical rule: if there is a fact you never want leaving your computer, do not save it as a memory. Memory is for things that are useful for the agent to know, and “useful for the agent to know” means “shared with the AI provider every time it might be relevant.” Anything you want kept strictly local belongs in your own notes, not in Clawless memory.

Settings you can adjust

A short tour of the controls. Open Settings from the gear icon at the bottom of the navigation rail, then click the Memory section in the sidebar on the left.

Auto-extraction on / off. Whether the classifier scans conversations for new memory candidates. On by default. Turning it off does not delete existing memories; it just stops new ones from being added automatically. You can still add memories manually with the classifier off.
Re-injection interval. How often Insights are re-fed into a long conversation to keep the agent grounded. Default is every ten messages. Lower means the agent stays grounded but you pay slightly more in tokens. Higher saves cost at the risk of the agent drifting from your earlier context.
Capacity-pruning notification. Whether to show a notification when General-tier memories are pruned to keep the database tidy. Off by default. Turn it on if you want visibility into what is dropping.
Decay rules. How long an unused memory takes to move down the tiers and eventually into Decayed. Defaults are reasonable. Advanced users can tighten or loosen them.

None of these settings are urgent. The defaults work for almost everyone. They are listed here so you know where to look if a particular behavior is surprising you.

Common first-day surprises

Five things people often notice about memory in the first week.

You did not type any of your Insights. They appear because the classifier read your conversations. Read through them once early on. Edit anything that is slightly off; pin anything that is load-bearing.
All your agents already know things you only told one of them. This is the shared-memory-pool default at work. Useful most of the time. If you want a specific agent that does not know anything about your other agents’ work, that is on the roadmap, not at launch.
A memory you have not thought about in months can come back. Decayed memories surface when something in the current conversation makes them look very relevant again. This is intentional. It is how the engine handles “I have not worked on this in a year, but here we go again.”
Editing a memory does not retroactively change past replies. The agent’s replies from yesterday were generated with yesterday’s memories. Today’s edited memory shapes today’s reply going forward, not the conversation history that is already written down.
The Pinned tier has a hard size cap. Around 2,200 characters total. Pinned content goes into every prompt, so the cap protects you from accidentally making every message slow and expensive. If you hit the cap, Clawless tells you and asks you to consolidate before adding more.

What you should remember

Memory is distilled facts; conversation history is the literal chat. Memory is compact and travels across conversations and agents; history is bulky and stays inside one chat. The Memory panel is where you see and control the memory side.
Four tiers, from most-important to least. Pinned is what you have explicitly told the engine to always remember (with a hard ~2,200 character cap). Insights are the engine’s high-confidence inferences. General is everything else, included only when relevant. Decayed is the path to “forgetting without deleting”: old memories that wake back up only if something current makes them look very relevant.
Three ways memories get in. The classifier picks them up automatically as you talk; you can ask an agent to remember something explicitly; you can type one directly into the Memory panel. Once in, you can edit, move, or delete any of them.
Memories live on your computer, but travel to the provider with each message. Anything you do not want a provider to see should not be saved as a memory.

In the practice that goes with this lesson, you open the Memory panel, review the Insights Clawless has gathered so far, pin one or two facts that should always travel with you, and add one manual memory. About fifteen minutes start to finish.