Scientific Foundation

Cortex implements 23 computational neuroscience mechanisms and 5 information retrieval techniques. Each is a faithful translation of published equations into working code — not a metaphor, not an analogy, not a "bio-inspired heuristic."

The Zetetic Standard

Every module follows a strict evidence protocol:

1. No source, no implementation. Every algorithm must trace to a published paper with exact equations.
2. Multiple sources required. A single paper is a hypothesis. Cross-reference before accepting.
3. Read the actual paper. Not blog posts, not summaries. The equations.
4. No invented constants. Every hardcoded number comes from paper equations, paper experiments, or measured ablation data.
5. Benchmark before commit. Every change is measured. No regression accepted.
6. Say "I don't know." A confident wrong answer destroys trust.

Where engineering defaults exist (signal weights, blend parameters), they are explicitly labeled in the code and accompanied by ablation data where available (see benchmarks/beam/ablation_results.json). They are never falsely attributed to a paper.

How to audit this codebase

Every module's docstring cites its source paper and the exact equations implemented:

grep -r "et al\." mcp_server/core/ --include="*.py" -l    # All paper citations
cat tasks/paper-implementation-audit.md                     # Full audit trail

Paper Index

Information Retrieval

Neuroscience — Encoding

Neuroscience — Consolidation

Neuroscience — Retrieval & Navigation

Neuroscience — Plasticity & Maintenance

Ablation Data

Ablation results are committed to the repository as JSON for reproducibility.

Rerank Alpha (Cross-Encoder Blend Weight)

Tested on BEAM 100K (20 conversations, 395 questions). Source: benchmarks/beam/ablation_results.json

Higher CE weight helps because conversational memory benefits more from semantic understanding (cross-encoder) than lexical matching (first-stage).

Documented Engineering Defaults

These values lack paper backing and are explicitly marked as such in code:

Architecture

Clean Architecture. Inner layers never import outer layers.

Storage: PostgreSQL 15+ with pgvector (HNSW) and pg_trgm. All retrieval in PL/pgSQL stored procedures.

Retrieval pipeline: Intent classification -> PG recall_memories() (5-signal TMM fusion) -> FlashRank cross-encoder reranking -> Titans surprise update.

Benchmarks

All scores are retrieval-only — no LLM reader in the evaluation loop. We measure whether retrieval places correct evidence in the top results. Nothing else.

Most memory systems report full QA scores (retrieve + GPT-4 reader + judge). This conflates retrieval quality with reader model strength. A strong reader compensates for broken retrieval. We don't do that.

Note: BEAM LIGHT comparison is full QA (LLM-as-judge from Table 2), not retrieval-only — shown for reference.

Known weaknesses: Abstention requires knowing what was never discussed — a comprehension problem, not retrieval. Instruction following requires surfacing meta-directives semantically distant from topical queries.

Development

pytest                    # 2068 tests
pytest tests_py/core/     # Core layer only

References

1. Ebbinghaus, H. (1885). Memory: A Contribution to Experimental Psychology.
2. Hebb, D.O. (1949). The Organization of Behavior. Wiley.
3. Collins, A.M. & Loftus, E.F. (1975). A spreading-activation theory of semantic processing. Psychological Review, 82(6).
4. Bienenstock, E.L., Cooper, L.N. & Munro, P.W. (1982). Theory for the development of neuron selectivity. J. Neuroscience, 2(1).
5. McClelland, J.L., McNaughton, B.L. & O'Reilly, R.C. (1995). Why there are complementary learning systems. Psychological Review, 102(3).
6. Abraham, W.C. & Bear, M.F. (1996). Metaplasticity. Trends in Neuroscience, 19(4).
7. Frey, U. & Morris, R.G.M. (1997). Synaptic tagging and long-term potentiation. Nature, 385.
8. Schultz, W. (1997). A neural substrate of prediction and reward. Science, 275(5306).
9. Bi, G.Q. & Poo, M.M. (1998). Synaptic modifications in cultured hippocampal neurons. J. Neuroscience, 18(24).
10. Kandel, E.R. (2001). The molecular biology of memory storage. Science, 294(5544).
11. Doya, K. (2002). Metalearning and neuromodulation. Neural Networks, 15(4-6).
12. Wixted, J.T. (2004). The psychology and neuroscience of forgetting. Ann. Rev. Psychology, 55.
13. Friston, K. (2005). A theory of cortical responses. Phil. Trans. R. Soc. B, 360(1456).
14. Hasselmo, M.E. (2005). What is the function of hippocampal theta rhythm? Hippocampus, 15(7).
15. Leutgeb, J.K. et al. (2007). Pattern separation in the dentate gyrus and CA3. Science, 315(5814).
16. Tse, D. et al. (2007). Schemas and memory consolidation. Science, 316(5821).
17. Turrigiano, G.G. (2008). The self-tuning neuron. Nature Reviews Neuroscience, 135(3).
18. Kanerva, P. (2009). Hyperdimensional computing. Cognitive Computation, 1(2).
19. Perea, G. et al. (2009). Tripartite synapses. Trends in Neuroscience, 32(8).
20. Yassa, M.A. & Stark, C.E.L. (2011). Pattern separation in the hippocampus. Trends in Neuroscience, 34(10).
21. Bastos, A.M. et al. (2012). Canonical microcircuits for predictive coding. Neuron, 76(4).
22. Dudai, Y. (2012). The restless engram. Ann. Rev. Neuroscience, 35.
23. De Pitta, M. et al. (2012). Computational quest for understanding astrocyte signaling. Front. Comp. Neuroscience, 6.
24. Sutskever, I. et al. (2013). On the importance of initialization and momentum in deep learning. ICML.
25. Kastellakis, G. et al. (2015). Synaptic clustering within dendrites. Prog. in Neurobiology, 126.
26. Buzsaki, G. (2015). Hippocampal sharp wave-ripple. Hippocampus, 25(10).
27. Kumaran, D. et al. (2016). What learning systems do intelligent agents need? Neurosci. & Biobehav. Rev., 68.
28. Gilboa, A. & Marlatte, H. (2017). Neurobiology of schemas. Trends in Cognitive Sciences, 21(8).
29. Stachenfeld, K.L. et al. (2017). The hippocampus as a predictive map. Nature Neuroscience, 20.
30. Nogueira, R. & Cho, K. (2019). Passage re-ranking with BERT. arXiv:1901.04085.
31. Josselyn, S.A. & Tonegawa, S. (2020). Memory engrams. Science, 367(6473).
32. Wang, C. et al. (2020). Microglia mediate forgetting via complement-dependent synaptic elimination. Science, 367(6478).
33. Ramsauer, H. et al. (2021). Hopfield networks is all you need. ICLR.
34. Bruch, S. et al. (2023). An analysis of fusion functions for hybrid retrieval. ACM TOIS, 42(1).
35. Wang, X. & Bhatt, S. (2024). Emotional modulation of memory. Psychological Review.
36. Maharana, A. et al. (2024). LoCoMo: Long context conversational memory. ACL.
37. Behrouz, A. et al. (2025). Titans: Learning to memorize at test time. arXiv:2501.00663.
38. Joren, D. et al. (2025). Sufficient context. ICLR.
39. Kosowski, A. et al. (2025). Dragon Hatchling: Memory management system. arXiv.
40. Wang, Y. et al. (2025). LongMemEval. ICLR.
41. Tavakoli, M. et al. (2026). BEAM: Beyond a million tokens. ICLR.

License

MIT

Citation

@software{cortex2026,
  title={Cortex: Scientifically-Grounded Memory System Based on Computational Neuroscience},
  author={Deust, Clement},
  year={2026},
  url={https://github.com/cdeust/Cortex}
}