Capture Surfaces

Capture Surfaces

Real life does not arrive as a clean document. It arrives through speech, meetings, handwriting, photographs, video, messages, and physical objects. Memova therefore treats capture as a family of surfaces that feed a common, user-owned memory layer.

The strategic principle is simple: capture should reduce the work required to preserve Context without weakening privacy, provenance, or control. A recording is useful only when the person can understand where it came from, correct the interpretation, and decide how long it should remain.

Memova Pen: A Handwriting System That Learns the Writer

Research supports the central technical hypothesis behind Memova Pen: writer adaptation and personalized handwriting recognition can materially improve accuracy. The right product architecture is not to train a new large model from scratch for every person. It is to begin with a general handwriting foundation model and maintain a lightweight writer profile, embedding, prompt, or adapter for each user.

That creates a more useful promise than “accurate OCR”:

The more you write, the better Memova understands your writing.

Offline Images and Online Trajectories

Handwriting recognition has two broad input modes.

Offline recognition sees the final image: a scan, photograph, screenshot, word crop, line image, or page image. It works with ordinary paper and has low hardware requirements, but it loses the writing process. The model must infer everything from the final shape, which is difficult for cursive writing, abbreviations, and highly personal styles.

Online recognition also captures the process: x/y coordinates, stroke order, timestamps, speed, pressure, tilt, and potentially IMU signals. These dynamics provide information that a static image cannot contain. For a smart pen that can obtain tip trajectories, online recognition should be substantially stronger than camera-only OCR in difficult personal handwriting.

The strongest architecture uses both. The rendered image describes what the glyph finally looks like; the trajectory describes how it was produced. A contextual language model then uses project names, people, calendar vocabulary, and nearby conversation to resolve ambiguity.

Evidence From Writer Adaptation Research

Several research directions support this design.

Meta-learning for a new writer shows that a system can adapt after observing only a small amount of handwriting from a previously unseen person. Reported gains vary by task and architecture, but studies have found meaningful performance improvements after a small calibration sample. This supports an onboarding experience in which a person writes one page and Memova begins learning their hand.

Prompt-based personalization, including work such as MetaWriter, treats personalization as prompt tuning rather than full-model fine-tuning. A frozen general model keeps its broad recognition ability while a very small set of writer-specific parameters adapts at test time. Self-supervised image reconstruction can make use of unlabeled samples, reducing the amount of correction the user must provide.

Explicit writer-style modeling adds a writer identity or style embedding to the recognition model. The model learns how one person forms characters and adjusts decoding accordingly. A product can retain a compact style representation without requiring the raw content of every note to leave the device.

Image and trajectory fusion combines a visual encoder with a trajectory model, often using early fusion. The image contributes glyph structure; the trajectory contributes direction, timing, and stroke sequence. Research on path signatures and pen-tip trajectories similarly indicates that dynamic features are central signals rather than optional metadata.

Multimodal and language models are useful as post-processing and context layers. They can rerank candidate text, restore names and terminology, correct spelling, and turn recognized writing into structured notes. They should not be the only recognizer: zero-shot performance varies widely across languages, historical documents, and complex handwriting, and a model’s self-correction is not always reliable.

Recommended Technical Architecture

Layer 1: Capture

The capture layer should retain:

• pen-tip x/y trajectories;
• timestamps and pen-up or pen-down state;
• speed and acceleration;
• pressure and tilt when available;
• IMU or camera data when required to reconstruct motion;
• a page image or rendered trajectory;
• optional audio Context from a meeting, classroom, or conversation.

For “write on any paper” hardware, the hardest engineering problem is reliable trajectory reconstruction. Sensor drift, sampling-rate alignment, surface movement, and ground-truth calibration must be solved before downstream models can perform consistently.

Layer 2: General Handwriting Model

A general model performs baseline recognition for a new user. Depending on the input, it may combine CNN, ViT, or Swin-style visual encoders with CTC or sequence-to-sequence decoding, plus a transformer over trajectory points.

The cold-start experience must already be useful. Personalization should improve a working system rather than rescue an unusable one.

Layer 3: Personal Writer Profile

Each user maintains a lightweight profile containing:

• a writer embedding or adapter;
• character-shape preferences;
• common abbreviations;
• a personal vocabulary;
• frequently used project names, company names, and people;
• correction history and confidence patterns.

The profile learns how this person writes particular letters and characters, how they connect strokes, which abbreviations they use, and which entities appear in their real work. Only a small fraction of the general model’s parameters should need to change.

Layer 4: Contextual Correction

An LLM or specialized language model can:

• correct spelling and spacing;
• identify people, projects, and proprietary terms;
• rerank multiple recognition candidates;
• use meeting audio to resolve ambiguous handwriting;
• reconstruct sentence order;
• transform raw notes into structured, searchable knowledge.

The model should receive candidate probabilities and relevant authorized Context, not invent a clean transcript without visibility into recognition uncertainty.

Layer 5: Feedback and Continuous Learning

Every correction is training evidence. If the recognizer outputs “metting” and the user changes it to “meeting,” the system can learn the person’s connected double-e, update vocabulary, and improve similar future cases.

Personalization progresses through three phases:

1. Cold start: use the general model with no special training requirement.
2. Implicit personalization: learn from ordinary corrections, accepted names, and recurring vocabulary.
3. Self-supervised adaptation: learn style through trajectory reconstruction, next-stroke prediction, masked image reconstruction, and other tasks that do not require manual labels.

Product Experience

An optional calibration Page can include letters, numbers, symbols, natural text, and the names that matter in the user’s work. The product promise can be expressed simply:

Write one Page. Let Memova learn your handwriting.

After onboarding, learning should happen through normal use. Correct a word once and the system remembers. Write a project name repeatedly and it joins the personal vocabulary. Finish a meeting note and the system can compare handwriting with the meeting’s audio Context.

The user should not have to manage a machine-learning workflow.

Privacy

Handwriting style and trajectory are sensitive behavioral or biometric data. The default should therefore be local-first:

• keep the writer profile on the user’s device whenever possible;
• do not upload raw trajectories by default;
• require explicit opt-in for cloud processing;
• support export and deletion of both raw samples and the learned profile;
• separate private recognition data from Pages selected for sharing.

This privacy architecture is not a compliance footnote. It is part of Memova’s product distinction: the system becomes more personal while the person retains ownership.

How to Evaluate the MVP

A useful experiment should recruit 20–50 people and collect 5–10 pages of real notes from each person: meetings, classes, task lists, brainstorming, mixed languages, and notes containing real names and project vocabulary.

Each sample should include a page image, tip trajectory, corrected ground truth, and—when authorized—contextual vocabulary. The comparison should include:

1. image-only OCR;
2. trajectory-only recognition;
3. image and trajectory fusion;
4. fusion plus a per-user adapter;
5. the personalized system plus contextual reranking.

Character Error Rate and Word Error Rate are useful, but product success depends on additional measures:

• error rate for people, companies, and project names;
• corrections required per 100 words;
• delay from writing to usable text;
• whether the user accepts the result without retyping it;
• whether action items and commitments survive recognition correctly.

The commercial value will not come from a small benchmark improvement alone. It will come from reducing the user’s correction burden and preserving the entities that make a note actionable.

Strategic Conclusion

The technical direction is feasible. Neither an LLM alone nor OCR alone solves personal handwriting recognition. The durable advantage comes from:

trajectory + image + personal adaptation + contextual language understanding

Livescribe and Neo Smartpen primarily answer, “What did I write?” Memova Pen should answer, “Why did I write this, who and what is it connected to, and what should happen next?”

The core product is therefore not a transcription device. It is a physical-world capture surface that turns handwriting, speech, and Context into actionable memory while becoming more accurate for one person over time.

Hardware Partnership Boundary

Partner hardware can accelerate industrial design, sensing, manufacturing, and distribution. Memova should retain ownership of the memory schema, the writer profile, context alignment, feedback writeback, privacy model, and the experience that turns capture into action.

The partnership claim must remain testable: define what the device captures, what the partner supplies, what Memova owns, how accuracy and latency will be evaluated, and what data is allowed to leave the user’s device.

Decision

Present Memova to ODMs as a Memory OS and agent-action system spanning five hardware entry points, not as a buyer seeking a quote for one commodity recorder.

Primary message:

One software system, five hardware entry points. Each device captures a different part of real-world context, while Memova turns that context into durable memory and user-confirmed action.

The five current consumer price directions are:

These are retail price directions, not target BOMs or final commitments.

ODM Added Value

Memova can help an ODM move from one-off manufacturing into reusable AI-native reference designs:

• software differentiation and potential ASP uplift;
• one customer relationship expanding across multiple SKUs;
• post-sale feature updates and recurring software value;
• privacy, consent, local-first processing, user confirmation, and secure data interfaces designed into the device;
• global professional-user positioning and a joint AI-hardware case study;
• aggregated, consented product feedback informing the next hardware generation.

The product distinction is not recording or transcription alone. Memova links captured input to people, projects, decisions, commitments, preferences, and next actions, then writes corrections and outputs back into memory.

External-Claim Guardrails

• “World first” may be used only for Memova's specifically defined Personal Superalignment platform category framing. Do not broaden it into claims that Memova invented personal AI memory, LLM Wiki, knowledge bases, or AI alignment.
• The earlier US$1.5M financing target is historical planning language and must be refreshed against completed financing before reuse.
• Founder revenue history, team roles, and academic credentials require approved public wording.
• Medical, legal, and accessibility use cases are research/target markets, not clinical or professional efficacy claims.