Academic foundations
We’ve tried to be honest about where ideas come from. If a feature has a paper, it’s listed here.
| Paper | Concept adopted | Where in the code |
|---|---|---|
| TMEM (arXiv:2606.04536) | Parametric memory via fast weight updates | recall/trust-decay.ts |
| MemQ (arXiv:2605.08374) | TD(λ) eligibility traces on provenance DAG | governance/td-lambda |
| RecMem (arXiv:2605.16045) | Recurrence-gated LLM invocation | recall/rerank.ts gating |
| HyperMem (arXiv:2604.08256) | Hypergraph memory with n-ary hyperedges | hyperedge |
| DELTAMEM (arXiv:2606.03083) | Residual trees for episodic delta-encoding | episode delta-encoding |
| UCE (arXiv:2606.02304) | Typed evolvable knowledge units + scheduler | the 10 typed primitives |
| Useful Memories Become Faulty (arXiv:2605.12978) | Consolidation trap → raw episodes as evidence | consolidation/cluster.ts gate |
| CALMem (arXiv:2605.20724) | Token-budget-adaptive injection depth | recall/render.ts |
| Experience Compression (arXiv:2604.15877) | Cross-level adaptive compression | summarisation |
| ACT-R (Anderson, 2004) | Activation-based memory retrieval | recall/search.ts strength |
| Hebb (1949) | Co-activation strengthening | mnemo_relation co-recall |
| STDP (Bi & Poo, 1998) | Spike-timing dependent plasticity | relation directionality |
Why this matters
Most “memory for agents” startups ship vibes. We ship citations and the tests to back them up. If a paper turns out to be wrong, we revisit the feature — but we don’t add features that lack a defensible source.
Adding new ones
PRs that introduce cognitive features must include:
- A citation to the paper or production system they derive from.
- An integration test that exercises the claimed behaviour.
- A benchmark line that shows the claimed savings.
The bar is high on purpose — the cognitive surface is where memory systems silently rot if you’re not careful.