DEV Community

Cover image for Take Control of Claude's Memory
Andrew Eddie
Andrew Eddie

Posted on

Take Control of Claude's Memory

Claude has memory. You probably turned it on recently when the web app asked and then... forgot about it.

Here's what you're missing.

Two Systems, One You're Ignoring

Claude actually runs two memory systems in parallel:

Auto-generated synthesis — Claude summarises your conversations and builds a
context profile automatically, updated every 24 hours. This is what most people
have passively enabled and never looked at.

Manual memory edits — You tell Claude exactly what to remember. Concise,
curated, yours.

Most people are only getting the first one. It captures everything: one-off
questions, situational context, noise. The second one is where the leverage is.

(Official docs on how both work)

The Limits (and Why They Matter)

  • 30 manual memory entries maximum
  • 500 characters per entry
  • Persists across every conversation

That's substantial curated context that future-Claude gets before you even say
hello — compared to starting cold, re-explaining your preferences, fighting
whatever's already in the context window.

The Problem With Passive Memory

Auto-generated memory captures everything with equal weight. What you actually
want is signal: how you think, what you care about, context that makes
every future conversation better rather than just the next one.

The goal isn't a fact sheet Claude glances at. It's a mental model that's
already activated when you arrive.

The Solution

At the end of any good conversation, use this prompt:

"Here's a conversation I want to mine for memory. Extract only what's
genuinely signal about how I think, what I care about, or context that would
make future conversations meaningfully better. Discard anything situational
or one-off. Propose entries in 500 chars or less, and flag if anything
overlaps or contradicts what you already have stored."

Claude will:

  1. Extract the signal
  2. Propose concise entries
  3. Check for overlap and contradiction
  4. Let you approve before adding anything

Compaction

As entries accumulate, treat them like git history: periodically squash them.
Merge overlapping entries. Promote patterns ("Andrew did this three times" →
"Andrew consistently does this"). Drop anything superseded by a better
formulation.

A simple prompt handles it:

"Review my current memory entries and propose a compacted version — merge
overlaps, promote patterns, remove anything redundant."

Why This Works

You're not training Claude (you can't). You're giving future-Claude better
priors. Instead of starting from zero, it starts knowing how you work, what you
value, how you approach problems.

It's like git for your AI relationship. Deliberate commits instead of auto-save.

One More Thing

Claude now supports importing and exporting your memory between AI tools —
useful if you want to migrate, back up, or seed a fresh start deliberately
rather than letting the auto-synthesiser decide what survived.

Try It

Next conversation that goes somewhere interesting, end with that prompt. See
what it extracts. Add what's useful. Compact when it gets noisy.

This is the natural next step from the Session Protocol
— that approach solved context within a project. This solves context across
your entire relationship with an AI assistant.

Your future self will thank you when Claude arrives already knowing who you are.

Top comments (0)