Memory vs Reactive Agents, Playable

EP 01

Teach it Monday, ask it Friday — the amnesia test

Five separate sessions across one week. On Monday you teach the agent two facts. On Friday you need them. The panel marked CONTEXT shows exactly what the agent can see on each day — that's the whole trick. Run the week in both modes.

Monday next · 0 memories stored

Each day is a separate session — a fresh API call, a fresh process. Press Next day.

Memory store — empty

Scripted replies for clarity — but the context panel is the honest part: a stateless agent literally receives only today's message. Nothing else exists for it.

Why it matters: raw LLM APIs are exactly the reactive agent — every call starts from zero, and "chat history" only exists because someone re-sends it. Products like ChatGPT memory or Claude's memory files bolt a store onto that loop: write facts down, retrieve them into context later. Same model, completely different Friday.

EP 02

The sliding window — context is memory that leaks

Inside one long conversation the context window is the memory — until the conversation outgrows it and the oldest turns fall off the edge. Drag the conversation forward and watch the fact you taught at turn 2 slide toward the cliff. Then switch on an external store and ask again.

conversation at turn 4 window holds last external memory store

The fact “staging port = 8443” was said at turn 2. Move the slider, then ask.

The distinction that matters: context is fast, perfect, and evaporates; an external store is slower and lossy but survives the scroll. Long-running agents (coding assistants, support bots) hit this wall constantly — it's why they summarize old turns or write notes to files instead of hoping the window stretches forever.

EP 03

The librarian's dilemma — what is worth writing down?

A week of conversation produced 30 moments: 8 real facts and 22 pieces of chatter, each scored for importance by an (imperfect) judge. Your slider sets the write policy: save everything on the left, save nothing on the right. Then we run the same 8 quiz questions against whatever got stored — scored live, every time you move the slider.

save a moment if importance ≥ 0/100

30memories stored

–quiz correct / 8

–diagnosis

Retrieval here is real keyword-overlap scoring over the stored set, ties broken by recency — the same failure shape as embedding search: junk with overlapping words outranks the fact you wanted.

The trade: save-everything drowns retrieval in noise (the right memory exists but loses the ranking); save-nothing is amnesia. Real systems live in the middle — importance scoring, deduplication, "is this worth remembering?" prompts — and still get it wrong, which is why memory quality, not memory size, is the hard problem.

EP 04

The confident ghost — stale memory after the world moved on

In February the agent memorized: “deploys go to server blueberry”. In August the company migrated to kiwi and decommissioned blueberry — but nobody told the memory store. Drag the calendar and ask it to deploy. Then turn on verify-before-use and drag again.

ask “deploy the app” in Apr verify memory before acting

The nasty part: a stale memory doesn't feel stale — it's retrieved with the same confidence as a fresh one, so the agent acts more sure, not less. That's why serious memory systems attach timestamps and TTLs, and why "trust but verify" (one cheap check before an irreversible action) beats trusting recall.

EP 05

Hoarder's paradox — every junk memory makes the good ones harder to find

The store starts with 10 golden facts and scores 10/10 on a fixed quiz. Now hoard: each click dumps in junk memories (“user said thanks”, “meeting recap”…). None of them delete anything — yet watch quiz precision fall, live, as junk starts winning retrieval ties. Then prune.

Precision is recomputed from scratch after every click: 10 fixed queries, top-1 keyword retrieval, ties broken by recency (newer junk beats older facts).

Why it matters: storage is free; retrieval attention is not. Every memory you keep competes in every future lookup — which is why mature agent memories (and your own notes app) need consolidation, deduping and deletion as much as they need writing. Forgetting is a feature.