continuous hypergraph memory for AI agents
Concepts, not chats. A local graph that accumulates mass on what gets ratified and decays what doesn't.
$ pip install bellamem copy
GitHub PyPI README Changelog AGPL-3.0

Live concept map click any viz to explore

These are real bellamem self-graph renders — the graph the tool built of its own development conversation, filtered to concepts with mass ≥ 0.7 (the ratified structural spine). Click a node for detail; drag, zoom, and explore. Same underlying graph, three different renderers.

Note: these renders are filtered to --min-mass 0.7 with turn hubs disabled. Only the ratified structural spine is visible — low-mass ephemerals and turn-preview text are excluded so nothing about this project's internal conversations leaks beyond concept topics.

The ontology every concept on two orthogonal axes

The ontology isn't decorative. It drives retrieval, decay, and dispute handling. A normative decision about testing style doesn't decay the way an ephemeral observation about yesterday's build does.

class — where does it come from, how is it used

invariant

hexagon · doesn't decay

time-stable principles and structural facts about the system

decision

diamond · revisable

commitments that constrain future action ("ship X before Y")

observation

ellipse · snapshot

a single empirical record about current state

ephemeral

rounded square · open→consumed

pinned plans with an explicit state machine and decay

nature — what kind of claim is this

metaphysical

amber · what the system IS

self-model, architectural facts, load-bearing invariants

normative

blue · what we commit to

rules, preferences, policies, code conventions

factual

green · measurable

checkable facts about the world, benchmark results

Try it locally it's ~30 seconds to your first concept map

pip install 'bellamem[viz3d]'

# Ingest a Claude Code session into your own graph
python -m bellamem.proto ingest /path/to/session.jsonl

# Open the concept map in your browser
python -m bellamem.proto viz --out v02.html --renderer 3d
xdg-open v02.html   # or `open v02.html` on macOS

Full README walks through the install, the class×nature ontology, the Before/After comparison, and the compression results across 15 real Claude Code projects.