plans/hackathon/04-the-memory-prize.md
Build an agentic tool that actually remembers.
Claude-Mem is the open-source memory layer for AI coding agents. A second agent watches your primary agent work, turns what happens into structured, timestamped observations, and makes them searchable — so the next session starts warm instead of cold. 100,000+ developers use it.
The mechanism, in plain English: your main agent takes actions — reads a file, edits a file, runs a command. Every one of those actions is handed to a second "observer" agent. The observer asks "is this worth a note?" Most of the time, no. When yes, it writes a structured note: a title, a subtitle, a few single-sentence facts (semantic chunks that stand alone), a short narrative, a category (bugfix, feature, decision, discovery…), some tags, the files involved, and a timestamp. The notes live in a local, meaning-aware search index. Next session, the titles come back as a timeline; details are fetched by ID only when needed. And the observer's job description — the mode — is swappable, so the same machinery can watch code, emails, photos, transcripts, robot logs, anything.
Build an agentic tool that actually remembers.
Not "an agent with a bigger context window." Not "an agent that stuffs everything into a prompt." An agent — or a tool for agents — that keeps what matters, finds it when it's relevant, and uses it to be faster, smarter, or more useful than it was the last time.
Claude-Mem gives you the memory layer. You build the thing that makes memory matter.
Pick one, combine a few, or bring your own.
An agent that opens with instant context instead of burning its first ten turns rediscovering the codebase. Measure it: how many turns did the cold version need before it was useful? How many does yours need? Show the difference.
Use the timeline and the memory-search skill as your retrieval layer for something new. The timeline is a time-ordered stream of structured notes with IDs — anchor on any one and read what happened around it. Search is semantic and keyword, filterable by project, category, date. What would you build if you had that as your database?
The skills in the repo are CLI-shaped: text in, text out. Wrap them in UI or UX that makes memory something you can see, steer, or share. A dashboard. A browser extension. A mobile view. A way to edit or approve notes. A way to hand your memory to a teammate.
Wire Claude-Mem into somewhere it doesn't live yet — your editor, your CI pipeline, your chat app, another agent framework or harness. Memory is most valuable where the work already happens.
Observations aren't text-only — Claude-Mem can take in images too, so screenshots, UI captures, design files and whiteboard photos are fair game alongside transcripts, logs, issues, commit history and support tickets. A custom mode tells the observer what to watch for. Any data in, any pattern out. /mode-creator will interview you and write the mode.
Observations land in real time while your agent works. Hook actions off them: a pattern shows up, something runs. A security note lands → ping the channel. A "decision" observation appears → open a ticket. The same error is observed three times → escalate. Agents that react to what just happened instead of waiting to be asked.
Use recall to cut tokens, turns, or wall-clock time. The titles-first recall pattern already reports ~99% savings in startup context; push it further, measure it, show the numbers.
Everything scores extra for being something someone would actually use.
A polished demo of a real need beats a clever demo of an imaginary one. If you'd keep using it Monday morning, you're on the right track.
$1,000 cash. Judged separately from overall 1st–3rd place — so you can win both.
npx claude-mem install, then use code FASTHACK30 at cmem.ai.