How AI agents remember: embeddings, decay, and a privacy list
Every agent platform says its agents "have memory." Almost none will tell you what that means. Here is ours, with the actual numbers: which embedding model, what the recall score is, how old memories fade, and the one list that overrides everything else.
Why a context window isn't memory
A model's context window holds the current conversation, and when it fills up, something has to give. Like every agent platform, we compact: an older stretch of conversation gets summarized so the work can continue. We'll say plainly what most vendors won't — compaction is lossy. A summary keeps what seemed important at the moment it was written, and discards detail you might need three days later. That is exactly why memory exists as a separate system: facts worth keeping shouldn't live or die with one conversation's context window.
So memory on Choreo is a store of small, durable facts — your name, your timezone, the project you keep coming back to, the format you like reports in — written deliberately, recalled by relevance, and visible to you the whole time.
The recall score, in full
When an agent needs to remember something relevant to what you just said, it doesn't reread every stored fact. Each memory is embedded with OpenAI's text-embedding-3-small (1,536 dimensions — fast and cheap, around $0.02 per million tokens), and recall is a scored search:
- Semantic similarity — the cosine similarity between your message's embedding and each stored memory's embedding. This is the backbone of the score.
- Times a temporal decay with a 180-day half-life — older memories are downweighted, but slowly: a three-month-old memory keeps about 70% of its similarity score, and a six-month-old keeps half. The curve is tuned so a strong semantic match always beats a weak recent one.
- Plus 0.15 for a key hit and 0.10 for a value hit — if your message literally contains the memory's label or text, that memory gets a bonus. Exact mentions shouldn't lose to fuzzy neighbors.
- And it fails open — if embedding is unavailable, recall falls back to keyword-and-recency search instead of failing. A degraded memory is better than none.
The retrieved set is small on purpose: the most recent facts united with the most relevant ones, deduplicated, capped — enough to inform a reply without drowning the conversation in everything the agent has ever been told.
remember, forget — and the list that overrides them both
Agents write memory through explicit tools, not by hoarding transcripts. remember stores one fact under a short label, with the value capped at 2,000 characters — a memory is a fact, not an archive. forget deletes by label. And every stored memory is visible and editable in your account's memories page, so "what does it know about me" is a page you can open, not a support ticket.
Above all of it sits the neverRemember list: patterns you configure that memory writes are checked against, silently. If you put your employer's name, a project codename, or any phrase on that list, an agent that tries to store a matching fact simply fails to — no error surfaced to the model to route around, nothing persisted. Agents are also instructed never to store secrets or credentials, and the instruction is backed by the same write path. You can watch your first memory get written in your very first conversation — they show up in the dashboard as they're stored, and the free plan needs no card.
What this doesn't solve
Memory is retrieval, not omniscience. An agent recalls what was deliberately stored and what scores well against the current message — it does not replay everything it has ever seen, and it can still be wrong. The honest framing: context windows forget by necessity, compaction forgets by design, and memory is the deliberate, inspectable layer that decides what deserves to survive both.
See it for yourself — free, $2 of credits once you confirm your email, no card.
Start free