$ cd ~/products/iris/memory
nf iris suite · 4/4
Iris Memory
Physics-Based Memory Graph
A neuroscience-inspired memory graph where concepts are nodes joined by Hebbian springs: co-activation strengthens a connection, time decays it, so strong associations survive and unused ones fade.
- +Hebbian springs — Co-activated concepts pull closer together
- +Natural decay — Unused associations fade instead of accumulating
- +Context recall — Continuity across sessions
The idea
Most assistant memory is a pile of stored text searched by similarity. Iris Memory is a graph that behaves more like a physical system. Concepts are nodes joined by springs. When two concepts come up together, the spring between them strengthens and shortens; when they don't, it decays. Strong associations survive and unused ones fade, instead of accumulating forever.
How it behaves
- Three timescales: short-term links decay over about a day, episodic ones over a week, semantic ones over a month
- Consolidation: every few turns, random walks over the graph reinforce the paths used most, much like rehearsal
- Structure: over time the graph settles into a small-world shape, tightly clustered, with short paths between distant ideas
- Stability: the system's energy only ever decreases, so the graph settles instead of oscillating
Recall
Recall combines two signals: similarity to the query, and activation spreading along the graph from related concepts. The second lets Iris bring back context that is associated with a question without resembling it word for word.
Research
The maths is written up in a paper that isn't published yet. The implementation is tested against it: the test suite checks the paper's results on convergence, decay, energy dissipation and small-world emergence directly.
Status
- [x]Graph dynamics, decay and consolidation implemented
- [x]Tests covering each of the paper's results
- [x]Persistence to SQLite
- [ ]Session memory
- [ ]Forgetting on request
- [ ]Retrieval benchmarks under load