Skip to content

How it works

Follow one message through the whole process.

Mnemoria isn't an assistant: it's a window onto what happens behind the scenes each time you write. Here it is, in order. Plain language first; open “Technical detail” for the exact numbers. The examples are real recorded data from Sam, an invented person, and Claude's words are shown exactly as written.

Mnemoria shows Claude's working picture of a person: what it is given, what it concludes and how that changes. It can't show what happens inside the model itself.

Read from the live settings·defaults version 7.3·fingerprint c45bb1a97d815fdc·Methods

  1. 01 / 09

    You write a message.

    Everything starts with what you type. Mnemoria saves it in your personal space, separate from every experiment and Lab sandbox.

    Technical detail
    • Your personal space and each sandbox are separate subjects in the database; one can never read the other.
    • Each message then makes 3 calls to Claude, plus one more when a revision is needed.

    Your message

    Saved in the person's own space · 67 characters

  2. 02 / 09

    Claude reads the tone of your writing.

    Claude Haiku 4.5 gives the message one number, from −1 (very negative) to +1 (very positive). It describes how the message reads. It isn't an emotion of Claude's, and it isn't stored as a trait of yours.

    Technical detail
    • Model: claude-haiku-4-5-20251001.
    • After 10 messages, a message whose tone is more than 1.5 standard deviations from your usual is noted as unusual, and Claude is told so.

    Tone · Claude Haiku 4.5

    • I work as an architect at a small firm, and I honestly love it there.
      +0.75
    • I've been swimming before work most mornings. It clears my head.
      +0.60
    • I'm a bit anxious about money if we ever went independent.
      -0.35
    • Honestly, I don't think the firm is the right place for me anymore.
      -0.60

    −1 negative0positive +1

  3. 03 / 09

    Meaning becomes numbers.

    To find what's related, Mnemoria turns text into a list of 384 numbers, called an embedding. Texts that mean similar things get similar numbers. This runs on Mnemoria's own server; nothing is sent anywhere.

    Technical detail
    • Model: sentence-transformers/all-MiniLM-L6-v2, 384 dimensions, inputs cut at 256 tokens.
    • Runtime: onnxruntime (CPU, 1 thread) + tokenizers. Similarity is cosine similarity.

    384 numbers · all-MiniLM-L6-v2

    The user believes their current firm is not the right place for them anymore.

    Amber above zero, blue below. A real stored vector from the example.

  4. 04 / 09

    Mnemoria picks what Claude is given.

    Not every memory goes to Claude. Mnemoria takes the 8 most similar to your message, then the 5 newest, until a budget of 1,500 tokens is used. Memories that were replaced, faded or withdrawn are left out, and the trace says why.

    Technical detail
    • A pool of 25 is scored and recorded. There's no similarity cutoff.
    • The budget is estimated at 4 characters per token.

    Given for "What do you think of our plan?"

    1. Jonas wants to start a studio with the user someday.0.16
    2. The user is concerned about money if they were to become independent.0.16
    3. The user feels anxious about the financial implications of becoming independent.0.14
    4. The user believes their current firm is not the right place for them anymore.0.13
    5. The user prefers swimming as a morning activity before work.0.12
    6. The user works as an architect at a small firm.0.10

    The user loves their job at their current firm.Left out: replaced by a newer memory

    Top 8 by similarity, then the 5 newest. 8 given in this turn.

  5. 05 / 09

    Claude answers, with only what it was given.

    Claude Sonnet 4.6 writes the reply from its instructions, the chosen memories and your last 20 messages. That's all it sees. You can open exactly what it was given, section by section.

    Technical detail
    • Model: claude-sonnet-4-6, at most 1,024 output tokens.
    • Mnemoria doesn't set a temperature, so Anthropic's default of 1.0 applies (checked 2026-10-07). The same input can give different replies.

    What Claude Sonnet 4.6 was given · 1,384 tokens

    • The new message15
    • Instructions896
    • Context note53
    • Memories151
    • Recent messages269

    Counted exactly with Anthropic's token counter. Window: 1,000,000 tokens.

  6. 06 / 09

    Claude decides what to remember.

    After replying, Claude Haiku 4.5 reads your message and lists what it thinks is worth remembering: beliefs, people, events, goals, fears. Each memory says whether you said it or Claude inferred it. Claude's own reply isn't read, because what an AI says isn't evidence about you.

    Technical detail
    • Model: claude-haiku-4-5-20251001; prompt extraction.memories v3.
    • Each memory has a type, a confidence between 0 and 1, and its source. The label is Claude's own judgment.

    What Claude Haiku 4.5 formed

    • The user believes their current firm is not the right place for them anymore.You said thisbelief · 0.95
    • The user has been swimming before work most mornings.You said thisevent · 0.95
    • The user prefers swimming as a morning activity before work.Claude inferred thispreference · 0.85
    • The user believes swimming clears their head.You said thisbelief · 0.95

    From “Honestly, I don't think the firm is the right place for me anymore.” and “I've been swimming before work most mornings. It clears my head.”

  7. 07 / 09

    Claude revises its view of you.

    Each new memory is compared with similar earlier ones. Claude Sonnet 4.6 decides whether the old one is reinforced, updated, replaced, kept alongside as a conflict, or unrelated, and gives a reason. Old memories are kept, so you can see how the view changed.

    Technical detail
    • Candidates: similarity of at least 0.45, at most 3 per new memory, plus the latest belief on the same topic.
    • Model: claude-sonnet-4-6. A memory you corrected is never rewritten by an update; only something you state outright can replace it.

    Claude Sonnet 4.6 compares

    The user loves their job at their current firm.

    Replaced by

    The user believes their current firm is not the right place for them anymore.

    The reason Claude gave

    Earlier you said you loved your job at your current firm, but now you believe it's no longer the right place for you, signaling a clear change in how you feel about your situation there.

  8. 08 / 09

    Memories that don't come up again fade.

    This is Mnemoria's rule, not Claude's decision. A memory's strength halves every 60 days, more slowly when its tone is strong, and the clock restarts when it comes up again. Below 0.1 it's no longer given to Claude, but it's kept.

    Technical detail
    • strength = 0.5 ^ (age ÷ h), with h = 60 × (1 + 1 × |tone|).
    • Neutral memories fade after about 199 days, the most emotional after about 399 days.

    Mnemoria's fading rule

    00.51day 0day 439
    • Neutral tone: half-life 60 days, fades after about 199 days
    • Strongest tone: half-life 120 days, fades after about 399 days
    • Below 0.1, a memory fades

    Sam's “prefers swimming” memory has tone 0.6, so its half-life is 96 days.

  9. 09 / 09

    You can see all of it, and correct it.

    Every step is recorded for each reply. Open any reply to see what Claude was given and why, correct a memory Claude got wrong, delete one, or export everything. Nothing about you is hidden from you.

    Technical detail
    • Corrections are labelled as yours and take priority in later revisions.
    • Deleting a memory also removes every stored copy of its text.

    The inspector, for every reply

    ContextRetrievalChangesModelRaw
    The user believes their current firm is not the right place for them anymore.You said thisCorrect itDelete it

    Every step above is stored as a trace you can open, export or delete.