Back to Claude Mem

Claude

plans/hackathon/claude-mem-cheatsheet.html

13.15.36.5 KB
Original Source

01 · The problem

AI agents wake up with amnesia

Every session starts from zero. That costs three things:

Without memoryWith Claude-Mem
Time — the agent burns its first turns re-reading the codebaseSession opens with a timeline of what happened last time
Money — re-reading = tokens = dollarsTitles-first recall — startup context typically reports ~99% savings
Lost decisions — "why did we do it this way?" Nobody wrote it downDecisions recorded with their rationale, searchable forever

02 · The idea

Two agents: one builds, one remembers

Main agent

The Builder

  • Reads, edits, runs commands, talks to you
  • Decides what to build
  • Never waits for the observer

Second agent

The Observer

  • Is handed every tool use + its result
  • Asks one question: "is this worth a note?"
  • Works silently in the background

It observes from outside. It never changes what the builder does. If it vanished, the builder would carry on — just with no memory afterward.

03 · Try it

You be the observer

Click an action the builder just took. Would you take a note?

📄 Reads a config file📂 Lists a directory✅ Tests go green after a fix🐞 Fixes a nasty bug⚖️ Picks SQLite over Postgres🏁 Session ends

👈 Pick an action

Every single tool use is handed to the observer. Most become nothing. A few become notes. This filter is the whole product.

04 · The unit of memory

Anatomy of a note

Every note has the same shape. Tap each field — the example note below updates as you explore.

#Title One line: what happened tap The most important field — most of the time the title is all the future agent reads. Write it so the title alone tells you whether you care. #Subtitle One more line of context tap Which part of the system, what the situation was. #Facts ~3 single-sentence bullets tap The semantic chunks. Each is a standalone true statement that makes sense with zero surrounding context — shaped to be found by search and dropped into a brand-new conversation. Test: could this sentence be pulled out alone and still be useful? If not, it's not a fact yet.

#Narrative A short paragraph: the story tap What was happening, what was tried, what was learned, why it matters. Facts are for skimming; the narrative is for understanding. #Category One label from the mode's list tap bugfix · feature · refactor · change · discovery · decision (in code mode). Answers "what kind of thing is this?" at a glance and makes the notebook filterable. #Tags Reusable knowledge labels tap how-it-works · why-it-exists · what-changed · problem-solution · gotcha · pattern · trade-off. Later you can ask: "show me every gotcha in this project." #Files + Time Where, and when tap Files read/modified tell the future agent where to jump. The timestamp is automatic — it's what makes the timeline possible.

● bugfix#1109623:48 PM

Fixed login redirect loop caused by stale session cookie

Auth middleware in the web app

  • The loop only happened when the session cookie was older than 24 hours.
  • The fix clears the cookie before re-issuing the login redirect.
  • The logic lives in the auth middleware, not the login page.

Looked like a frontend routing bug; the cause was server-side. Clearing the expired cookie before redirecting broke the loop.

tags: problem-solution · gotcha · files: middleware/auth.ts

05 · Getting notes back

Recall, cheapest first

TIMELINE Automatic at session start. Titles + IDs + category icons only. Fifty notes cost a few hundred tokens.

GET BY ID On demand. A title looks relevant → fetch that note's facts + narrative. Pay only for what you need.

SEARCH For everything older. By meaning, keyword, category, date, project. Titles come back first — then fetch details.

Mantra: titles first, details on demand. Never dump everything.

06 · The superpower

Modes: swap the observer's job description

A mode is a plain config file: who the observer is, what the categories are, what the tags are, what language to write in (30+ shipped). Pick a mode to see what changes:

Code (default)Code — ChillEmail investigationLaw studyRobot monitoring✨ Your mode

Code Development

Watches a coding session: files read, edits, commands, test runs.

The default. Records what got fixed, built, decided, and discovered — with rationale.

Don't write one by hand — run /mode-creator. It interviews you, then writes, installs, and activates the mode. Any data in, any pattern out.

07 · Do this today

Five-minute quickstart

npx claude-mem install — a note-taker joins the room.

Work in a session. The observer starts taking notes immediately. (Open the local viewer to watch them land live.)

Start a second session in the same project → the timeline appears at the top. That's memory. First session seeds; second is where you feel it.

Ask "did we already fix X?" — watch the agent search, filter titles, fetch by ID.

Run /mode-creator and point the observer at something that isn't code — a photo, a transcript, a chat export.

The challenge

Build an agentic tool that actually remembers

$1,000

cash · judged separately from overall 1st–3rd — you can win both

  1. Warm boot — instant context, no ten-turn rediscovery
  2. Build on the timeline — timeline + search as your retrieval layer
  3. Give the skills a face — UI/UX so memory can be seen, steered, shared
  4. Build an integration — editor, CI, chat app, another agent framework
  5. Ingest anything — images, transcripts, logs, tickets + a custom mode
  6. Fire on what it sees — react to observations in real time
  7. Memory as a speed play — fewer tokens, turns, seconds

Tiebreaker: would someone actually use it Monday morning?

Every hacker gets 30 days of CMEM Pro free: npx claude-mem install, then code FASTHACK30 at cmem.ai.

08 · Speak the language

Vocabulary

Observerthe second agent that takes notesObservationone note (title, facts, narrative, category, tags, files, time)Tool useone action the main agent takes — read, edit, run, searchTimelinethe time-ordered list of note titles injected at session startModethe observer's job description — role, categories, tags, languageSkilla small reusable helper command (text in → text out)Session summarythe end-of-session "where we are / what's next" noteWarm bootstarting a session already knowing what happened

Repo & skills: github.com/thedotmack/claude-mem 30 days Pro free: code FASTHACK30 at cmem.ai