Personal World Models

Hana Azab · María Benavente Kinship Technologies

Paper (PDF) · DOI · Project page · Package · Supplement

Memory is a compression problem. Current systems treat it as retrieval, which can say what happened but not what it means or what comes next. A Personal World Model (PWM) is a life compressed the way a map compresses territory: what recurs becomes structure, and any model can read the structure and anticipate from it. GOLGI instantiates the framework: it ingests a real four-week photo archive — 720K tokens of raw life — into anchored entities, routines, and narratives an agent traverses in hundreds of tokens, on a pipeline that runs end-to-end on 4B on-device models.

Results

An agent reading the committed structure, against a matched flat-retrieval control and production memory systems over the same archive (50 personal-history questions, fixed answer and judge models — only the memory layer varies):

Memory layer	Correctness
GOLGI	46
GraphRAG	30
Supermemory	29
Matched flat-retrieval control	19
Flat retrieval (vanilla)	17
Mem0	15
Zep	6

• Structure is causal, not decorative: corrupting the structure degrades answers in step with its fidelity (r = 0.81 across 36 corrupted variants, permutation p < 10⁻³); random groupings of the same shape give the agent almost nothing.
• Anticipation: predicting from the places the structure anchors retrieves the exact masked, unseen moment 32% of the time where a structure-free baseline reaches 0% (chance < 1%); replicates across a second encoder (DINOv2) and a second subject.
• On-device: 4B models (Ollama, Q4_K_M) reach up to 82% end-to-end parity with the cloud reference across all four pipeline capabilities.

The pipeline

Photos enter as Moments; Primitives (faces, objects, activities, spaces, text) are extracted per moment; recurring Primitives are promoted to Anchors with user confirmation; the anchored stream is synthesized into Narratives — Plans, Trips, Routines, and Chronicles. Everything is computed once, at ingestion, and read by agents through MCP.

Try it on your own camera roll

The evaluation archive is private (it is a life); the intended reproduction is to run the pipeline on your own photos. The package (macOS arm64, Python 3.12) ingests a photo directory end-to-end in cloud (Gemini) or fully on-device (Ollama) mode and exposes the structure over MCP:

pip install golgi-0.1.0-*.whl
golgi ingest -p my_persona -m on-device   # or -m cloud
golgi mcp -p my_persona --transport stdio

The supplement contains the four ingestion-stage prompts verbatim, the judge rubrics, model/quantization/decoding configurations, promotion thresholds, and the benchmark harness with all baseline adapters.

Citation

@misc{azab2026pwm,
  title  = {Personal World Models},
  author = {Azab, Hana and Benavente, Mar{\'i}a},
  year   = {2026},
  doi    = {10.5281/zenodo.21655904},
  url    = {https://personalworldmodels.github.io}
}