Bee: Building an Ambient Agent That Remembers and Acts

Personal context without the homework

A reflective account of your day is useful even if you never write a journal. An assistant that remembers your preferences saves you from explaining them again. Bee starts with these everyday needs: capture context as life happens, turn it into memories, and use those memories to help.

In the February 2025 Latent Space conversation, Bee combines an audio wearable with transcription, speaker attribution, and personal memory processing. Worn in a wristband or as a clip, the device has microphones and a mute button. Recall and to-do tracking provide immediate uses; anticipating needs is the larger ambition.

The idea predates the wearable. A guest recalls starting a personal AI company in 2016 with a co-founder named Esther. Its chat app asked questions and offered feedback, trying to understand a person who changes over time. Two constraints held it back: teaching the system about yourself took substantial effort, and the available language technology struggled to turn that information into useful dialogue. Some younger users loved answering questions; most people did not want the homework.

That venture pivoted to Squad, a product for watching video with friends, and was sold to Twitter. The guests worked together on Twitter Spaces and other products before returning to personal AI shortly before ChatGPT launched. They tried an app again, then moved toward ambient capture—collecting context without requiring users to open an app and enter it.

Why use a separate wearable?

A phone already has microphones, but continuous capture competes with its other jobs. Calls interrupt audio acquisition, and users must remember to restart it. For an assistant intended to remain present throughout the day, those small interactions become substantial barriers.

Apple Watch support let existing owners try Bee without buying another device. The team worked on background operation and battery consumption, but its watchOS implementation still required activation and restarting after calls, even with reminder notifications. Some Watch users reportedly bought Bee hardware after a few days; the team attributed that interest partly to less restarting and charging friction, without reporting a conversion rate.

The daily summary can provide a reward before a large personal history accumulates. One speaker who did not keep a journal found the end-of-day notification unexpectedly interesting: it connected events they had just lived through. That reflective value could appear after a single day.

Dedicated hardware also gives Bee more control over audio acquisition. The team said the Apple Watch framework available to it did not allow the desired gain and sample-rate adjustments. Better input helps, but the models still need to cope with environments as different as a quiet studio and a noisy restaurant; tuning for one can degrade another. Bee therefore works with the phone, using the wearable as the assistant’s ears.

From continuous audio to usable memories

The app shows a continuously updating account of the current conversation alongside summaries of previous days. Producing that account requires more than transcription. Speaker identification helps prevent someone else’s statement from becoming a fact about the wearer. Users can teach Bee familiar voices to improve attribution and its understanding of relationships.

Conversation endpointing determines where one conversation ends and another begins. Those boundaries are fuzzy: one interaction can flow into the next. Bee uses voice activity, semantic content, and other signals to divide the day into logical units. Once a conversation ends, larger models revisit it to improve the interpretation and extract a summary, key points, tone, and possible action items. A daily walkthrough brings events and locations together.

These records support recall and reflection while also feeding an evolving model of the person. Bee exposes the resulting personal facts for inspection, keeping the user involved in what the system believes about them.

In the recall demonstration, a question about a Taiwan manufacturing trip produces trip details and manufacturing problems, with references back to the underlying conversations. Bee can also search the web and consult Gmail and Google Calendar. An answer can therefore combine a spoken discussion, an email, and a calendar entry rather than depend on a single transcript passage.

One speaker estimates their accumulated context at roughly 50 million tokens and daily information output at about 200,000 tokens, without specifying the counting method or included material. A more playful request asks Bee for a Spotify Wrapped-style review of life in 2024. It generates travel and shipping statistics alongside reflections—an example of the experience, not verified life metrics.

A notification becomes an action

A beta WhatsApp integration extends recall into proactive help. It evaluates incoming notifications against two criteria: whether they are important enough to surface and whether Bee can help. A co-founder’s request for an Italian restaurant recommendation passes both.

The demonstration separates suggestion, authorization, and execution: 1. Bee offers to find a nearby restaurant and send a recommendation. 2. The user authorizes it through chat or by holding the wearable’s multipurpose button and speaking. 3. An Android cloud phone, with access to the user’s accounts, searches for restaurants, opens the recipient’s WhatsApp thread, and sends a response. 4. Bee reports completion without requiring the user to watch the execution.

Bee reports sending Maria a recommendation for Bottega with rating and distance information. The ensuing discussion distinguishes task completion from recommendation quality: one participant considers the restaurant acceptable but criticizes its pizza and likes some of its pastas. A sent message does not establish that the assistant understood the diner’s taste.

Bee also supports customizable wake words; one participant uses “Hey, Alfred.” Programmability offers another way to make the assistant fit individual needs, particularly when users already manage their lives in other applications.

At recording time, Bee’s API was read-only, with action support planned as those capabilities became more generally available. Real-time output through a socket connection could already feed external systems that performed actions. Developers could connect that output to their own automation; invoking Bee’s action system through the API remained a future capability.

Attention, changing minds, and privacy

Persistent memory raises questions about relationships. Could conversation records become instant replay for an argument? How should the system handle conflicting public and private statements? One response prioritizes helping users interpret emotionally charged interactions. The speaker sees an AI without the user’s emotions as an objective perspective—a product belief, rather than a measured property of Bee’s judgments.

Proactivity adds the problem of when to interrupt. Frequent suggestions become annoying; excessive restraint misses useful opportunities. The team expects personal context to help, but choosing the right moment still requires strong reasoning. Stored context also needs maintenance as people change their minds. Users can currently confirm, reject, or edit facts; automating that process remains a goal.

Bee processes audio without retaining it by default and centers summaries on the wearer. Speaker identification, processing, prompts, and storage choices are intended to limit personal information retained about bystanders. Whether transient audio processing followed by summarization constitutes recording remains legally unsettled in the discussion; it establishes no exemption from consent requirements. The team also treats bystander privacy as an ethical concern even where recording is permitted.

Acceptance varies. Two speakers say nobody had asked them to switch Bee off, but users had reported uncomfortable partners and workplace policies that they believed prohibited wearing it around sensitive work. Features being added included geofencing, to disable Bee at specified locations, and concept fencing, to exclude specified topics. Their effectiveness was not demonstrated.

Professional demand was an unexpected early use. The team reported shipping more than 100 devices daily and attracting people who talk for a living. One user reportedly asked Bee to review job interviews and suggest improvements. That illustrates a coaching use, without demonstrating better interview outcomes. Bee’s intended role is an understanding engine supporting specialized applications built by others; the team says the eventual consumer killer feature remains unknown.

Hardware people will keep wearing

Bee approached CES through a three-day visit with planned partner and media meetings, without a company booth. The team wanted people to experience the product and reported achieving its awareness goals. Its practical advice was to arrive knowing whom to meet: access to media and suppliers was valuable, but the volume of simultaneous launches made attention hard to win.

One supplier meeting prompted an exploration of body-heat harvesting. Bee drew more power than that approach could supply. A solar cell covering the device’s face seemed more plausible, while kinetic harvesting was reportedly insufficient. These remained possibilities, not implemented features.

Wearability had already forced revisions. One participant removed the original pendant and forgot to put it back on. Community feedback pushed Bee toward a bracelet after roughly three or four hardware iterations: the circular pendant was bulky, and some women said it conflicted with necklaces and other accessories. The resulting modular design lets the device move between wristband and clip.

The target is a device users can wear and forget. That favors low weight, modest on-device computation, and long intervals between charges. The team describes battery life improving from approximately 35 hours in an earlier version to almost seven days, without specifying a test workload. A phone-mounted MagSafe design could hold a larger battery, but would leave the sensor wherever the phone happened to be—including a pocket.

These constraints also explain why Bee launched without vision despite building an early visual prototype. The team identifies radio transmission as the main power bottleneck for sending images, even at low frame rates. Continuous all-day operation would require an unacceptably large battery. Placement adds a separate problem: a wrist camera is impractical, while a chest camera may miss what interests the wearer. Head-mounted capture better addresses that mismatch, but makes appearance central to adoption. Images also introduce further privacy challenges.

Manufacturing changes the iteration cycle

One team member came from computer science with little firmware experience, supported by a colleague with deep hardware expertise. Prototyping felt increasingly approachable, with Sonnet helping engineers get started on firmware. Manufacturing brought the larger jump: procurement, regulations, component sourcing, tooling, and cost.

Enclosure tooling creates long commitments. Bee submits its design for externally made production tooling and describes a four-to-six-week wait. A tooling problem can block enclosure production, requiring decisions and planning much earlier than a software team might expect.

Bee’s first circuit-board fabrication and assembly happened in Los Angeles for proof-of-concept work and field testing. Fast domestic small batches were possible at a high price; fabrication and assembly later moved to Taiwan. The team’s supplier-selection advice was concrete: - Ask other founders for references. - Match the supplier’s specialty to the work, such as plastics or circuit boards. - Visit in person to assess the partner and establish a relationship. The team was happy with the Taiwan process, but emphasized how stressful supplier selection and component sourcing become under tight deadlines.

Spend inference where it matters

Processing an entire day raises a different budget question. The team argues that falling token prices make large text inputs less daunting, without supplying a measured per-user cost. Speech-to-text remains the harder part because Bee combines real-time processing with a later pass through a larger model.

The first saving is voice activity detection: check whether speech is present before transcribing. The team describes detection as cheap and much of a typical day as containing no voice activity. Model requirements then vary by task. In its assessment, summarization needs less capability than a Sonnet-level model, while agent execution needs that higher level. Bee also self-hosts models, though it does not identify which models or stages.

The existing pipeline requires tedious tuning and trial and error. A team member proposes replacing more of it with an end-to-end model that learns the task together, potentially using a large language model for transcription. That could initially require more compute; distillation is suggested as a possible route to reducing the cost. The prediction that conventional speech-recognition systems will soon become obsolete remains a forecast.

The pricing discussion pairs inexpensive hardware with inference described as free, but the team plans subscriptions for capabilities such as agent execution. Detailed usage limits and per-user costs are unspecified, leaving the service’s long-term economics open.

Memory must handle time, uncertainty, and repair

Bee built its memory infrastructure internally because retrieval depended heavily on its data. The team considered general frameworks useful for getting started, but extensive customization favored a bespoke system. A simple temporal example shows why: a plan to visit London should stop describing a current plan once the traveler returns.

Bee describes its custom retrieval system as using small models for massively parallel retrieval, without detailing the implementation or giving a quantitative comparison. Separately, its person model distinguishes user-confirmed facts, treated as ground truth, from uncertain inferences. Confirmed context helps assess whether a fuzzier inference is plausible; removing the confirmation step remains an aspiration.

Implicit preferences require more than collecting explicit statements. A participant repeatedly orders the same dish from a favorite restaurant without announcing that preference. Their desired assistant would recognize the pattern and reorder it when asked. The team’s response emphasizes giving reasoning enough time to examine receipts, emails, and conversations together. In its experience, a key-value store with retrieval layered on top is insufficient for a growing corpus containing stale memories and possibly hallucinated source material.

Knowledge graphs offer another way to represent the data, but Bee does not use them extensively. The reported obstacle is making graph input understandable to the language model at inference time. Useful storage still requires selective retrieval and formatting that avoids overwhelming or confusing the model.

The core reliability requirement is recoverable, correctable memory. Even perfect transcription leaves missing context, and visual input would not eliminate every misunderstanding. Incorrect interpretations must be repairable so they do not undermine the accumulated personal model.

When personal agents exchange context

The closing experiments extend memory across people. Two agents reportedly used their users’ shared preference for French food and their presence in Pacific Heights to suggest a restaurant neither had visited. The participants say they went, describing it as a toy use case. Calendar integration could add availability to such negotiations. A separate gift experiment resulted in candles.

That introduces an unresolved authorization question: what may a personal agent disclose, and to whom? Proposed controls include permission scopes specifying which personal context can be shared and human sign-off. Whether every disclosure should require approval remains open. Opt-in location sharing provides an analogy; small automatic updates to parents offer another possible use.

Users had also asked whether Bee could assess compatibility with another person. The team had considered the idea but was reluctant to build a dating product itself. A personality-assessment anecdote captures the appeal and uncertainty of inferred self-knowledge: one speaker asked Bee for a Big Five personality assessment, received a characterization as disagreeable, and disagreed. It was an amusing reflection, not an accuracy evaluation.

At the close, the device was offered for $49.99 through the Bee website, and the team was hiring AI engineers who were scrappy and adaptable. Its closing prediction was that always-on AI would expand through startups and incumbents alike, with its effects on everyday life still to be discovered.