It is RAG. Here is what we extract, and how we check it.

The critique of agent memory is that most of it is retrieval with extra steps. Fair. The steps are the product, so here they are: what is extracted, how it is redacted, how it is stored, how it is served to the model, and how it is graded.

The engram is a store of nodes about one person, each a claim extracted from that person's own words, with a type, the stream it came from, the source it cites, and the dates it applies. The decision call retrieves the nodes that bear on a situation, with the precedents. A separate call applies the person's voice to the wording of what goes out.

Consent first: the owner chooses each source stream in writing. Extraction reads only what the owner wrote: sent mail, review comments, recorded decisions and corrections, transcripts and chat where connected. Mail received is parsed only as the thread a reply answers. Notifications and automated mail are excluded.

Redaction happens at ingestion, before storage: named people, companies and addresses become role types; no message body is stored; every node cites its source identifier and the character span it came from. The store is Postgres first, with valid_from and valid_to on every node, so a reversed rule supersedes the old one instead of overwriting it. A temporal graph database can come later if it earns its place; today its gains over those two columns are vendor numbers.

Then a persona card, used only to voice outbound text, and a fidelity run against held-out decisions. Each is a section below.

One extracted claim, as the store holds it. Types are rule, decision, preference, playbook and voice. The source is an identifier and a span, never a body. Roles stand where names were. The dates say when the claim began to apply and, if it was superseded, when it stopped and by what.

{
  "id": "n01",
  "type": "rule",
  "claim": "Never past 10% on reorders. Volume earns terms, not margin.",
  "stream": "sent-mail",
  "source": {
    "message_id": "<redacted-at-ingestion@example>",
    "span": [
      13,
      86
    ]
  },
  "roles": [
    "a buyer at a restaurant group"
  ],
  "valid_from": "2025-03-12",
  "valid_to": null,
  "superseded_by": null
}

The decision call gets the situation, the retrieved nodes and the precedents, and nothing about personality. A separate call applies the persona card to the wording of outbound text. The reason is measured, not stylistic: persona text in a prompt can lower accuracy. "Principled Personas" (EMNLP 2025, nine models, twenty-seven tasks) found expert personas roughly neutral and irrelevant persona detail costing nearly thirty points; later studies found expert personas underperforming no persona on factual work. One paper dissents and argues those nulls come from flawed methods. So we run the test rather than argue it: agreement with and without the card, and the card stays out of decisions if it lowers agreement.

Corrections become nodes. Consolidation runs between sessions, in the background, the way sleep-time agents review recent conversations and update memory without interrupting the work. The owner reviews the proposed nodes each week.

No model is tuned on the owner, and that is a choice, not a limit. Open-weight tuning is cheap enough that a rival can truthfully claim a model tuned on its customer. We decline it for three reasons. A deleted node is gone from the next answer, and every node can be read; weights cannot be read or edited line by line. Retrieval stays current more cheaply than retraining. And the store moves with the owner to whichever model fits, so the model is a part we can swap and the engram is the part that is theirs.

At least three hundred of the owner's own decisions are held out, each later in time than every node in the store, with any thread that mentions a held-out decision removed from the corpus. They are stratified by decision type and by source stream. Each situation is replayed to the helper, and agreement is counted.

The ceiling is the owner's own consistency: we re-ask a sample of decisions later and report agreement as a fraction of that test-retest rate, with its interval. At three hundred decisions a seventy-five percent rate carries a ninety-five percent interval of about plus or minus five points. The design follows the strongest individual-level study we know: agents built from two-hour interviews with 1,052 people reached 83 to 86 percent of each person's own two-week consistency on held-out survey answers, in the version of 28 June 2026.

The score is reported per decision type, and voice is scored separately, because the research says tone is the weak dimension: in-context imitation of over four hundred real authors approximated news and email reasonably and struggled with informal registers, and a 2026 benchmark of personal twins found persona tone and memory recall the weakest of six capabilities.

The published scores for agent-memory systems are vendor-run and disputed. One vendor reports above ninety on a long-conversation benchmark; a rival says the comparison misconfigured its system and that the benchmark's conversations fit inside a modern context window, where a plain full-context baseline wins. The rival has a stake, and the critique is specific and checkable. Our position is simpler than taking a side: a leaderboard score says nothing about whether a helper does what you would do. So we grade on your own decisions, and publish the method before any number.

Tone and casual chat are the weakest dimensions, so the helper drafts and the owner approves. Fidelity decays over a long conversation, so each decision card starts fresh from the nodes rather than from the thread. Sycophancy induced by long-term memory has barely been studied; one benchmark from February 2026 states it had not been evaluated before. A thin record gives a wide interval, and we say so rather than publish a number the sample cannot carry.

An engineering engram is built from review comments, corrections and recorded decisions, scoped per repository: what you asked for, what you blocked, and why. It becomes the rules a coding agent loads before it works, and it is graded by replaying past reviews: would the agent have blocked what you blocked? The founder's own engram counts code review comments among its richest streams.

We route each task to the model that fits, through API terms that exclude training on your data, and we name no vendor on this site. The engram is the part that is yours. The model is a part we can swap.