Tang
Co-founder
Senior Data and Analytics Engineer
Seattle · Hong Kong
faren发轫
Research and development · Last updated September 4, 2026
Building systems that help important work keep moving.
Faren is a research and development practice exploring continuity: how intelligent systems can preserve the state of interrupted work, maintain its context, and make it easier to resume.
The research begins with a working system we use ourselves.
01the question
What if setting a project down did not mean losing its momentum?
People rarely work on one important thing at a time. Priorities change, dependencies stall, and attention moves elsewhere. The real cost of returning is not the pause itself; it is reconstructing what mattered, what changed, and what should happen next.[1][2][8][9][10]
This is not a problem limited to one kind of brain. Neurodivergent and neurotypical people may manage attention, memory, and transitions differently; the system should support that range without treating one pattern as the default.[11]
02the current experiment
Faren is our attempt to make continuity a property of the system rather than another burden on memory. We treat it as four functions, each answering one way that paused work fails.
A · CAPTURE
Preserve the state
Record the decisions, dependencies, and unfinished work surrounding a project, not only the files and windows that were open.[17]
B · MAINTAIN
Keep context current
Organize what changes while attention is elsewhere and identify when the work becomes relevant again.
C · DECIDE
Choose, on the record
03the system
The current research separates two related problems: forming useful context from a person's record, and adapting a model to reason consistently about continuity.
flow runs top to bottom · click a layer to expand · hover a status to isolate it
The sources a person already generates: project files and notes, calendars, sleep and activity metrics from consumer wearables, and connected tools. Each source is enabled individually. Health means sleep, activity and recovery; no clinical data is ingested.
signals, with source and time
Signals are normalized to a common schema, timestamped, and stored with provenance: source, time of capture, and any transformation applied. The state record stores derived facts, not raw messages; raw signals are kept only as long as provenance needs them.
splits into two stores
A small, explicit, human-readable record of what a project needs to move: decisions taken, open dependencies and their status, unfinished items, and the intended next action. This is the unit the research tests first.
The larger body of notes, documents and research material, chunked and embedded for retrieval. Kept separate from the state record so the effect of structure can be measured apart from the effect of retrieval.
state and evidence rejoin
Watches incoming signals for events that bear on paused projects (a dependency resolved, a deadline moved, a decision made elsewhere) and writes candidate staleness events to the record for the person to confirm or reject. The most speculative layer, and the one that separates a continuity system from a resumption-note tool. Thread 2, Study M1, runs last.
a record kept current while attention is elsewhere
At inference the model is conditioned on passages retrieved from the record and corpus rather than on memorized content, so readout evidence is current by construction and every claim can be traced to the record that supports it.[3]
the evidence for one project, at the moment of return
Readouts are generated today by a hosted model pinned to one version, through interfaces whose terms exclude use of the data for training. Every readout stores its model version, prompt hash and retrieved-passage identifiers, so the record stays auditable when the model changes.
The planned base model runs locally. It is introduced in shadow mode, generating a logged but undisplayed readout for every case, so paired outputs can be compared blind before display switches over. Adapting it (supervised fine-tuning, then preference tuning[4][5][7]) happens only against a measured failure of the simpler baselines.
a readout, with its model version and evidence attached
A dataset of real resumption cases is being built from daily use. Retrieval, generation and interface behavior are tested as separate components and end to end, with human review as the reference and automated judges checked against it.[12] Study R1 first, then D1, C1, M1.
what passed review
Four required parts: the project's state, the evidence for each element of it, a recommended decision among resume, hold and abandon, and a next action. Each line opens to the record behind it. See section 04 for an example.
Status as of September 2026; yellow marks what runs today. Retrieval supplies current evidence without retraining the model.[3] Readouts are generated today by a hosted model pinned to one version; the planned base model is open-weight and local, and adapting it (supervised fine-tuning, then preference tuning[4][5][7]) is conditional on a measured failure of the simpler baselines. Every readout stores its model version, prompt, and retrieved evidence so the record stays auditable across that transition.[12] Full architecture and evaluation program in Working Paper 01 →
04a faren readout
ready to resume
faren research readout · 04
Portfolio site redesign
Paused for
3 weeks
Where you stopped
The new homepage layout is complete. The project-page template remains unfinished.
What changed
The photography has arrived. No other dependencies are outstanding.
Recommended next step
Build the first project-page template.
Resume
Hold
Abandon
05research principles
Operational validation
Hypotheses are evaluated in a working system against defined tasks, failure cases, and measurable outcomes.
Data governance
Personal data remains user-controlled. Local processing is prioritized when it is compatible with the system requirements.
Traceable outputs
Each recommendation should preserve an evidence path: retrieved material, applicable rules, model version, and known uncertainty.
Modular evaluation
Retrieval, model adaptation, inference, and interface behavior are tested as separate components as well as an integrated system.
Cognitive variability
Evaluation should include different patterns of attention, memory, and task transition rather than assume a single cognitive baseline.
Evidence discipline
Primary sources only. Record the actual numbers and where they came from, check replication, record the criticism alongside the finding, and mark confidence honestly.
06the working paper
Working Paper 01 sets out the framework: continuity as a system property, decomposed into four functions (capture, maintain, decide, resume), each with a hypothesis that can fail and a planned study; the testbed architecture with implemented and proposed layers distinguished; a comparison of existing tool categories; the evaluation program; and the popular claims we checked and do not use.
Spence-Chan, J., and Spence-Chan, T. (2026). Continuity systems: a conceptual framework and research agenda for intelligent support of work in progress. Faren Research Working Paper 01.
Faren is currently a working system, used and evaluated in day-to-day practice. Its next form will be shaped by the evidence rather than a predetermined roadmap. We will revise this record as the research develops.
07the people
Faren is an independent collaboration between Tang and James, partners and co-founders bringing complementary experience in data engineering, architecture, operations, and applied AI.
James
Co-founder
COO at Ensemble · Architect, AIA
Seattle · Hong Kong
08the name
发轫 — fā rèn — describes the moment a wheel is released and begins to move.
For us, it is a useful idea: momentum does not require everything to move at once. It requires important work to remain capable of moving forward.
09working bibliography
A starting set of references informing the current hypotheses and technical direction. The full bibliography, with the search method and the claims we checked and do not use, is in Working Paper 01. Inclusion indicates relevance to the work; it does not imply that every method has been implemented or independently replicated by Faren.