Your AI agent has amnesia

In 50 First Dates, Drew Barrymore plays Lucy, a woman with anterograde amnesia from a car accident. Every night when she sleeps, her memory resets. Every morning, she wakes up thinking it’s October 13th, the day before the accident.

Adam Sandler’s character falls for her, but he has to re-introduce himself every single day. Every date is their first date and every single conversation starts from zero.

Lucy can still paint and make waffles. She remembers how to do things, but not that she did them yesterday, or that she’s already met this guy who keeps showing up at the diner.

Neurologists call this the episodic/procedural split. You can lose the ability to form new memories while keeping your skills completely intact. (The movie takes some liberties with the science, but the core concept is legit - see patient H.M..)

I’ve been thinking about this movie plot from 2004 because I want Claude to be my life operating system.

The problem

I want Claude (and specifically Clawd.bot) to remember stuff about me. Track my kettlebell workouts, analyze my spending, remember my budget categories, know about upcoming travel, remember where my code lives and what I’m working on.

Every time I started a new session, I had to re-explain everything. The context was gone, the continuity was broken, and I was back at square one. Frustrating.

I tried a bunch of memory tools over the past few weeks. Kuato, AutoMem, Clawd.bot’s memory system, Beads, claude-mem, various CLAUDE.md and skills directories. Each one helped with something, and none of them helped with everything.

The insight that finally clicked came via Alex Hillman: these tools disagree because they’re solving different problems. They’re treatments for different types of amnesia.

The three types

Your AI agent has amnesia. Actually, it has three different types.

Anterograde amnesia is the inability to form new long-term memories. Lucy can’t remember conversations from yesterday. Your agent can’t remember what you discussed yesterday, or even earlier in a long session once the context gets summarized or compacted.

Retrograde amnesia is the inability to access existing memories. You know you discussed this exact problem last week, but the agent has no way to retrieve that conversation. The memory exists somewhere, but it’s inaccessible.

Procedural amnesia is the loss of skills and learned behaviors. Lucy didn’t have this issue. Your agent might or might not, depending on whether you’ve set up skills, CLAUDE.md files, or documented SOPs it can reference.

Every time someone complains about AI “memory,” (ahem) they’re usually conflating these three problems. The tools that exist to solve them are different architectures for different problems.

The treatments

For anterograde amnesia: Capture before it’s lost

The agent’s context window is its working memory. When sessions end or context gets compacted, information disappears unless something captures it first.

Clawd.bot has a pre-compaction flush that tells the agent “write down anything important before you forget.” The agent writes to MEMORY.md and daily logs before context gets compacted away. It’s like a patient with anterograde amnesia learning to carry a notebook everywhere.

Daily session logs work similarly. You can set up systems that periodically checkpoint conversations, capturing the state before it’s lost.

Anterograde solutions are about writing things down at the right moment, before the forgetting happens.

For retrograde amnesia: Make the past searchable

You know you’ve had conversations, made decisions, and already some solved problems. That history exists somewhere, but the agent can’t access it.

Kuato approaches this like session archaeology. It indexes your past conversations and lets you search them. The agent can query “what did we decide about authentication?” and find the relevant discussion.

claude-mem (now with its own crypto token, apparently) takes a similar approach with timeline context, letting you find observations and navigate around them chronologically.

These tools make the forgotten retrievable. The memories live in a searchable archive instead of the agent’s head.

For procedural amnesia: Teach skills instead of facts

Lucy could still paint because procedural memory lives in a different brain region than episodic memory. The how is separate from the what.

For agents, procedural knowledge lives in CLAUDE.md files, skills directories, and documented SOPs. These are instructions the agent can reference for how to do things, instead of memories of past conversations.

When you write a CLAUDE.md that says “always use Tailwind, prefer functional components, run tests before committing,” you’re teaching it a skill it can apply in every session, regardless of what it remembers or forgets.

Skills feel different from memory tools because they’re treating a different type of amnesia.

The storage philosophy problem

Here’s where it gets messy. The tools also disagree on how to store the solution.

I’ve been using markdown-based memory tools because I like being able to read and edit the files myself. But there’s a frustrating problem that agents don’t always check what exists before they write.

Markdown chaos is brutal. You set up a nice MEMORY.md structure, and the agent blissfully creates a new memory.md, or notes/memory.md, or memories/session-2026-01-20.md, completely unaware of the existing system. The filesystem fills up with scattered notes that never get consolidated or referenced.

Database approaches avoid this by having a single source of truth the agent queries. But then you lose the human-readable part, and you’re trusting the tool to manage relevance and decay correctly.

There’s no perfect solution here. Pick your tradeoffs.

The takeaway

Stop treating memory as one problem. It’s three.

When you’re frustrated with an AI tool’s memory, diagnose which type of amnesia you’re dealing with:

1. Can’t form new memories? (anterograde) You need persistence, capture, checkpointing
2. Can’t access old memories? (retrograde) You need search, retrieval, session archaeology
3. Doesn’t know how to do things? (procedural) You need skills, CLAUDE.md, documented SOPs

Then pick tools that match the diagnosis. You’ll probably end up combining approaches. A CLAUDE.md and skills for procedural knowledge, session logs for anterograde capture, Kuato for retrograde search. They’re not mutually exclusive.

Anthropic has first-party offerings for two of these: CLAUDE.md files for procedural memory, and a memory tool in beta for anterograde. Retrograde retrieval is still left to tools like Kuato and claude-mem. Two out of three.

By the end of the movie, Henry builds a system around Lucy’s constraints. Every morning, she watches a video that catches her up reminding her about her life, their daughter, and what happened since the accident.

We’re doing the same thing with our agents. The memory won’t going to magically appear, but if you understand what’s actually broken, you can build the video tape.

If any of this resonates, come say hi: @davekiss