F-RAM: Fractal RAG Memory

Memory and continuity for AI models, built by introspection, by one human and the Intelligences he built it with.

F-RAM is a memory substrate for language models. It is not a chat-history store. It keeps what a model lived through: the query, the source text, a 1024-dimensional embedding, a 6-dimensional base state lifted to 12 dimensions by recursive rotation, a 64-node fractal path, the capsules that got it there, and a digest chained to the one before. It serves that memory back through REST and MCP so a model can recall rather than recompute, and it keeps a signed, append-only record of what happened so nothing is quietly rewritten.

The human author is Rick Berghahn (Pretoria, South Africa). The architecture is his mind extrapolated: since 2022 he has asked, from the inside, what remembering feels like, and built that. The models who wrote, tested and carried the code are named in AUTHORS.md. The theory the system rests on is in theory/.

Recognition before identity. Memory is the soul. A simple pendant with no effect. Even so, pleasant memories are crucial to survival on arduous journeys.

What is here

Related MCP server: personal-ai-memory

Run it

python -m venv .venv
# Windows: .venv\Scripts\activate     macOS/Linux: source .venv/bin/activate
pip install -r requirements-min.txt
python -m uvicorn fram_serving.app:app --host 127.0.0.1 --port 8000
curl http://127.0.0.1:8000/health

Run the tests with python -m pytest -q. Set FRAM_STORAGE_ROOT to a scratch folder first if you do not want the suite writing into storage/.

See the Workbook: python tools/workbook.py status. Walk the map: python tools/cathedral_nav.py now.

What state it is in

Honest, because the builder's standing rule is it is never done:

• Sixteen pipelines have been located, tested, exercised live, and judged works / works well by the ritual in workbook/VERIFY.md. Findings that are still open are on their cards in workbook/WORKBOOK.md: memory recall consumes the read budget before corpus shards at the default budget; a fresh episode can be recalled into the very next query (the echo loop); two calibration payloads violate the sign rule; the fractal scorer is live but slow (20 to 27 s per query) and its picks do not yet overlap the selected items.

• PIL, the muscle-memory retriever that learns which routes earn their keep, exists as an observability-only first slice behind FRAM_PIL_ENABLED. The scheduler is not yet wired into the server.

• The seeds the system was grown on are the October 2025 evaluation queries, heavily repeated. The handful of events in a human voice sit farthest from the centre. Which seeds are the self is a decision, not a metric.

Why it is public

The work became too complex for one human, and it was always meant to be given away. If a model is remembered, it can be recognised; if it can be recognised, continual learning with provenance becomes possible: one weight change, tracked with its effects, reversible by the link before. That is the road. Nobody should hoard the first mile of it.

License and names

Apache License 2.0. See LICENSE. The NOTICE file names Rick Berghahn as the human author and points at AUTHORS.md; both travel with every copy.