Agents in Amnesia (AIAIO) – the game that teaches you about how you use AI

Notifications You must be signed in to change notification settings

Fork 0

Star 4

BranchesTags

Open more actions menu

Folders and files

NameName

Last commit message

Last commit date

Latest commit

History

74 Commits

74 Commits

.claude

.claude

.github/workflows

.github/workflows

docs

docs

examples

examples

mockups

mockups

qa

qa

scripts

scripts

skills/aiaio

skills/aiaio

src

src

tests

tests

.gitignore

.gitignore

AGENTS.md

AGENTS.md

CHANGELOG.md

CHANGELOG.md

CLAUDE.md

CLAUDE.md

LICENSE

LICENSE

PRIOR-ART.md

PRIOR-ART.md

README.md

README.md

index.html

index.html

package-lock.json

package-lock.json

package.json

package.json

tsconfig.json

tsconfig.json

vite.config.ts

vite.config.ts

Repository files navigation

Operation: Inner Space, except the world is your agent's actual session log.

🎮 Play the hosted demo (**THE OPENCLAW

HERMES CAMPAIGN**, 12 explicitly fictional levels). The real game is your own history: clone this repo, run npm run scan, and every Claude Code, OpenClaw, or Hermes session on your machine becomes a playable level. Your prompts are the tasks, your errors are the monsters, your compactions are the wall chasing you.

Beta notes. Desktop and keyboard required (mobile gets an honest gate). SessionCards are generated locally and never leave your machine. Sharing a card with someone is possible but experimental: cards contain redacted snippets of your real prompts, so read public/cards/.json yourself before you send it to anyone. A proper pre-share review UI is the headline of the next beta.

Your real session is the level. Its timeline is the terrain you cross. Your real errors spawn as monsters at the points where they actually happened. Your real tasks sit in the world as work stations. And behind you, always, the WALL OF FORGETTING (context pressure made spatial) advances, eating terrain, tasks, and eventually you.

Reach process exit 0 alive. Clear your task queue on the way for a perfect run; in your real-history campaign, exiting after recovering at least half of the recorded task work earns progression credit. Every token you spend fighting is distance the wall gains.

If you like this game or have ideas, hit me up on Twitter or Instagram @bradmillscan.

Run it

npm install npm run scan # optional: auto-build levels from YOUR sessions (see below) npm run dev # open the printed URL npm run build # static bundle in dist/

Production builds make no network calls and ship only the fictional public campaign. Local enrichment is an explicit opt-in dev action and may call the AI command you configure after showing its privacy notice.

Or let your agent do all of it

This game is made for people who run agents, so the agent can be the installer, the level curator, and the tech support. Hermes users:

hermes skills install https://raw.githubusercontent.com/sene1337/aiaio/main/skills/aiaio/SKILL.md

then tell your agent things like "set up aiaio", "find my 10 most dramatic sessions and make them levels", or "make the announcer a bitter golf commentator". Any other agent (Claude Code, OpenClaw, Codex): point it at AGENTS.md — same playbook, no skill required. If the vault comes up empty, npm run doctor prints exactly what was scanned and why each file was rejected; your agent can read it and fix the cause.

Campaign presentation is customizable: agents can write titles, display copy, and Observer lines over an immutable source-card snapshot. Factual campaigns retain source-derived mechanics. An explicit Remix request may select one of the fixed gentle, balanced, or brutal combat profiles; it has separate progress and is always labelled Remix.

Play

←/→ move. ↑ jump. ↓ fast-fall (explore drilled craters). space fire (toward facing).

W (hold) work the task at a station. You're rooted and heads-down while working, which means +25% damage taken. Working is how you win, and working is when you're weakest.

U open a ⬆ crate's command menu (choose with 1/2/3), or install a ◈ model upgrade. C runs /compact.

[ ] / 1-9 switch weapons. Slot 1 is the ∞ print-debug zapper: you can never run out of print statements, and you can never win with them alone.

S spawn a subagent (900tk plus ~16tk/s upkeep, because inference isn't free and the drip literally speeds up the wall). It's a lower model: weak zaps, cap of 2, loyal right up until a hallucination-ghost touches one or it falls into the wall. Then it's corrupted: red, garbled, and shooting at you. Terminate it or outrun it (rogues OOM-kill themselves after about 18s).

M mute. All audio is synthesized WebAudio (chip/glitch, zero assets): compaction is a stuttering descent into static, the wall has a heartbeat when it's close, tasks chime when they ship.

The systems

The wall of forgetting is action-driven. It does NOT creep on a timer. Every token you burn becomes wall distance (0.16px per token): firing, working, subagent upkeep, damage spew, even walking (reading the transcript is inference too, at 1tk per 10px). Stand perfectly still and it stands still with you. Your token bill is the storm. overflow-emitters spam tokens into your meter while you're near, which moves the wall, because everything does. Tasks it passes are forgotten (garbled, unrecoverable). It eats subagents whole. You it merely ruins: inside the zone you bleed 9hp/s and your weapons spray wildly, though you can still dive in, since unclaimed crates in there still work. Involuntary compaction (crossing your threshold) surges it 240px or more, plus the amnesia: shield gone, progress rewound, and from your 3rd compaction even completed tasks can un-ship.

/compact (C). Voluntary, anywhere, 20s cooldown. It drains your meter cleanly, with no surge and nothing forgotten, but the summarization pass costs 250tk and roots you heads-down for a beat. Run it too close to the threshold and the pass itself tips you over. Compact early, compact often.

Context economy. Every token burned is wall distance, and the meter itself is your countdown to the next involuntary compaction surge. The context-window-nuke erases half a screen of errors and floods a quarter of your own meter (about 700px of wall), so choose violence carefully. Getting hit injects error-spew into your context (stack traces are long), so damage accelerates your own compaction.

Upgrades. ⬆ patch crates open a 3-option command menu: the classic random gamble is always option 1, plus two seeded picks from /restore cached-subagent (no upkeep drip), /tune context-manager (+5% threshold), /patch shield-buffer, and /restock error-log. The rare ◈ MODEL UPGRADE crate stays guaranteed-good: more context budget, a higher threshold, a shield refill, and a vN title-bar tick.

Music. A generative ambient score (pure WebAudio): a warm pad and pentatonic plucks when you're safe, morphing darker, sparser, and more detuned as the wall closes, down to drone and static inside the forgetting.

Handicap. Low-stability sessions (yours was rough) grant a starting shield and a damage bonus, with the "why" line quoting your real numbers.

The bestiary (spawned from your error log)

Your error category Becomes Behavior

timeout ⏱ timeout-blob tanky lobber; shots detonate late

hallucination 👻 hallucination-ghost phases, teleports, is sure it exists; your hardening resists its touch

regression ☢ regression-splitter splits into two minis on death

restart / crash 🔁 restart-crawler relaunches itself once after dying

false_positive ⚡ false-positive-sniper telegraphed laser, 100% confidence, ~70% accuracy

tool_error 🔧 tool-turret interrupt bolts; hit while working = lose task progress

context_overflow 📈 overflow-emitter priority target: spams tokens into your meter while you're near, which moves the wall

recovery ➕ recovery-sprite friendly, touch for hp/shield

Enemy count per category scales with the error count, and weapon ammo scales from the same log, so your worst error category is both your biggest threat and your deepest magazine.

The semantic layer: the level tells your session's story

When a card carries them (the extractor mines all of this automatically):

goal. Your first real ask, shown as the mission: "the mission, in your own words: …"

Tasks are your actual messages. When logs have no structured tasks, your substantive user asks become the task stations, named in your words, placed at their real timeline positions, with work-units proportional to how much of the session each ask actually consumed.

Enemies spawn where the errors really happened (errors[].at).

◇ moments. Real wins ("ok it works. nobody touch anything.") and frustrations ("still broken. still. broken.") stand in the world where they happened. Walk past one and the transcript quotes it.

Compaction garbles your own words. The banner corrupts real lines from the session instead of canned filler, and the recap tells you where in the real session your run ended.

All of it is heuristic, deterministic, and on-machine, with no LLM involved. Redaction applies to every snippet, but a card still contains fragments of your actual prompts, so skim before sharing.

Campaign enrichment: an explicit local action

The main menu's ✦ ENRICH YOUR HISTORY offers two honest shapes:

SHAPE MY OPENING requires six eligible sessions and preserves chronological order.

BUILD MY CAMPAIGN requires fifteen and curates 15–24 eligible sessions.

It displays a consent notice before it sends redacted SessionCard excerpts to the configured AI command. No session is automatically enriched merely by opening or playing it. The raw card remains unchanged; enrichment publishes a versioned CampaignManifest overlay with source digests, ordered entries, and presentation copy. If the writer is unavailable, AIAIO atomically publishes a complete deterministic baseline campaign instead.

For an agent-driven flow, use the same engine:

npm run enrich -- --selection hardest --pace intense --remix brutal

default writer is claude -p; set AIAIO_LLM_CMD for another CLI agent

--remix gentle|balanced|brutal is explicit. gentle uses 70% hostile budget, 80% damage, +10 shield, and no Observer injections; balanced preserves the normal values; brutal uses 150% hostile budget (cap 36 total / 10 per type), 125% damage, and removes the stability handicap. Remix unlocks never alter factual-history ranks.

npm run scan: your sessions become levels

Auto-discovers agent session logs (~/.claude/projects, ~/.openclaw including archived agents, and ~/.hermes, whose SQLite history is dumped to JSONL automatically, or any root you pass), builds a SessionCard per session, and writes them to public/cards/, where THE MEMORY MAP makes a factual chronological campaign and LIBRARY keeps every session searchable and playable. --all scans the entire archive. The default takes the 12 most recent per root.

Quality gate. Only sessions with real extractable human asks become levels. Cron jobs, heartbeats, and ask-less machine runs are excluded entirely, because the game never fakes personalization. Selecting or starting a level never asks an agent to enrich it; enrichment only runs from the visible command or an explicit agent/CLI request.

Read-and-aggregate only. Log content is inert data, never executed or followed.

Samples are redacted (API keys, tokens, JWTs, credentials, emails, hex blobs) and truncated to 80 chars, but skim public/cards/*.json before sharing.

public/cards/ is gitignored, so your session data never lands in the repo.

You can also drop any SessionCard .json on the menu, or build one by hand: node scripts/extract-sessioncard.mjs -o card.json.

QA telemetry (dev mode)

Playing via npm run dev records gameplay telemetry to qa-logs/*.jsonl (gitignored): every fire, work, damage, compaction, and death event, first-use timing per control, and a snapshot every 2s (position,

[truncated for AI cost control]