$
Research & Citations
shodh-memory is grounded in decades of cognitive psychology and neuroscience research. Every constant in our codebase has a citation. Here are the papers that shaped our architecture.
📄 Read the Paper (PDF)27
Papers cited
10
Research areas
200+
Tuned constants
$
Cite shodh-memory
If you use shodh-memory in your research, please cite it using one of the formats below.
BibTeX
@software{sharma2026shodh,
author = {Sharma, Varun},
title = {shodh-memory: Cognitive Memory for AI Agents},
year = {2026},
url = {https://github.com/varun29ankuS/shodh-memory},
doi = {10.5281/zenodo.18668709},
version = {0.1.90},
license = {Apache-2.0},
note = {Hebbian learning, 3-tier architecture (Cowan 2001), hybrid decay (Wixted 2004), spreading activation (Anderson 1984)},
}APA 7th
Sharma, V. (2026). shodh-memory: Cognitive memory for AI agents (Version 0.1.90) [Computer software]. https://doi.org/10.5281/zenodo.18668709
IEEE
V. Sharma, "shodh-memory: Cognitive memory for AI agents," version 0.1.90, 2026. doi: 10.5281/zenodo.18668709
[01]
Memory Decay & Forgetting
How memories fade over time and why power-law decay matters
memory-decay-forgetting.viz
strength
100% |████
80% |██████
60% | █████▓▓
40% | ███▓▓▓▒▒
20% | ▓▓▒▒▒░░░░
10% | ▒░░░░░░░░░
+─────────────────────
0 1d 3d 7d 30d
exponential → power-lawOn the Form of Forgetting
1991
Wixted, J.T. & Ebbesen, E.B.
Psychological Science, 2(6), 409-415
DOI: 10.1111/j.1467-9280.1991.tb00175.xPsychological Science, 2(6), 409-415
How we use it: Foundation for our hybrid decay model. Demonstrated that forgetting follows power-law, not exponential, over long periods.
The Psychology and Neuroscience of Forgetting
2004
Wixted, J.T.
Annual Review of Psychology, 55, 235-269
DOI: 10.1146/annurev.psych.55.090902.141555Annual Review of Psychology, 55, 235-269
How we use it: Comprehensive review linking psychological forgetting curves to neural consolidation processes. Informed our 3-day crossover point.
Reflections of the Environment in Memory
1991
Anderson, J.R. & Schooler, L.J.
Psychological Science, 2(6), 396-408
DOI: 10.1111/j.1467-9280.1991.tb00174.xPsychological Science, 2(6), 396-408
How we use it: Showed that memory decay mirrors environmental statistics. Justifies our adaptive decay rates.
[02]
Hebbian Learning & Synaptic Plasticity
Fire together, wire together—the basis of associative memory
hebbian-learning-synaptic-plasticity.viz
before: after: ○ ─── ○ ● ═══ ● | | | | ○ ─── ○ ● ═══ ● weak edges strengthened (0.10) (+0.025/co-access) co-activation → stronger bonds
Synaptic Modifications in Cultured Hippocampal Neurons
1998
Bi, G.Q. & Poo, M.M.
Journal of Neuroscience, 18(24), 10464-10472
DOI: 10.1523/JNEUROSCI.18-24-10464.1998Journal of Neuroscience, 18(24), 10464-10472
How we use it: Measured actual synaptic strengthening rates (~3-7% per activation). Calibrated our HEBBIAN_BOOST_HELPFUL constant.
The Organization of Behavior
1949
Hebb, D.O.
New York: Wiley
New York: Wiley
How we use it: The foundational work on associative learning. 'Cells that fire together, wire together' is the core principle of our knowledge graph.
[03]
Spreading Activation
How activation spreads through associative networks
spreading-activation.viz
○
/ \
○ ○ ← hop 2 (0.49)
/ \ \
● ● ○ ← hop 1 (0.70)
\ /
★ ← query node (1.00)
activation = weight × 0.7^hops
surfaces related contextSpread of Activation
1984
Anderson, J.R. & Pirolli, P.L.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 10(4), 791-798
DOI: 10.1037/0278-7393.10.4.791Journal of Experimental Psychology: Learning, Memory, and Cognition, 10(4), 791-798
How we use it: Defined the mathematical model for spreading activation. Our SPREADING_DECAY_RATE and hop limits come from this work.
spreadr: An R package for simulating spreading activation in a network
2019
Siew, C.S.Q.
Behavior Research Methods, 51, 910-929
DOI: 10.3758/s13428-018-1186-5Behavior Research Methods, 51, 910-929
How we use it: Modern implementation of spreading activation. Validated our importance-weighted decay approach.
[04]
Sleep & Memory Consolidation
How replay during rest strengthens memory traces
sleep-memory-consolidation.viz
awake consolidation stable ░░░▒░░▒░ → ▒▓▓▓▓▒▒░ → ████▓▒░░ fragile replay durable traces cycles engrams hippocampal replay → cortical storage (our maintenance cycles mirror this)
About Sleep's Role in Memory
2013
Rasch, B. & Born, J.
Physiological Reviews, 93(2), 681-766
DOI: 10.1152/physrev.00032.2012Physiological Reviews, 93(2), 681-766
How we use it: Comprehensive review of hippocampal replay during sleep. Inspired our memory replay maintenance cycles.
Cognitive Neuroscience of Emotional Memory
2006
LaBar, K.S. & Cabeza, R.
Nature Reviews Neuroscience, 7, 54-64
DOI: 10.1038/nrn1825Nature Reviews Neuroscience, 7, 54-64
How we use it: How emotional arousal enhances memory. Our REPLAY_AROUSAL_THRESHOLD prioritizes emotional memories.
[05]
Interference & Competition
How similar memories compete and interfere
interference-competition.viz
memory A [███████▓▓▓] sim=0.92
memory B [████████▓▒] sim=0.95
│
▼
retrieval competition: B wins
A suppressed by interference
threshold > 0.85 → inhibition
(prevents redundant recall)Critical Issues in Interference Theory
1973
Postman, L. & Underwood, B.J.
Memory & Cognition, 1, 19-40
DOI: 10.3758/BF03198064Memory & Cognition, 1, 19-40
How we use it: Defined retroactive and proactive interference. Our INTERFERENCE_SIMILARITY_THRESHOLD detects competing memories.
Interference and Inhibition in Memory Retrieval
1996
Anderson, M.C. & Neely, J.H.
In E.L. Bjork & R.A. Bjork (Eds.), Memory (pp. 237-313). Academic Press
In E.L. Bjork & R.A. Bjork (Eds.), Memory (pp. 237-313). Academic Press
How we use it: Explained inhibitory mechanisms in retrieval. Informed our competition factor for similar memories.
[06]
Cognitive Architecture
Computational models of human memory
cognitive-architecture.viz
+---------------------------+
| +---------------------+ |
| | +---------------+ | |
| | | Working | | | <- seconds
| | +---------------+ | |
| | Session | | <- hours
| +---------------------+ |
| Long-Term Memory | <- permanent
+---------------------------+
Cowan's embedded processesEvolving Conceptions of Memory Storage, Selective Attention, and Their Mutual Constraints
1988
Cowan, N.
Psychological Bulletin, 104(2), 163-191
DOI: 10.1037/0033-2909.104.2.163Psychological Bulletin, 104(2), 163-191
How we use it: The foundational paper for our 3-tier architecture. Cowan's model distinguishes activated long-term memory from the focus of attention—our Working → Session → LongTerm tiers map directly to this.
The Magical Number 4 in Short-Term Memory
2001
Cowan, N.
Behavioral and Brain Sciences, 24(1), 87-114
DOI: 10.1017/S0140525X01003922Behavioral and Brain Sciences, 24(1), 87-114
How we use it: Working memory capacity limits. Refined our understanding of tier boundaries and why memories must be consolidated.
How Can the Human Mind Occur in the Physical Universe?
2007
Anderson, J.R.
Oxford University Press
DOI: 10.1093/acprof:oso/9780195324259.001.0001Oxford University Press
How we use it: ACT-R cognitive architecture. Our base-level activation and decay equations derive from ACT-R.
Memory—a Century of Consolidation
2000
McGaugh, J.L.
Science, 287(5451), 248-251
DOI: 10.1126/science.287.5451.248Science, 287(5451), 248-251
How we use it: Landmark review of memory consolidation. Our TIER_PROMOTION_WORKING_AGE_SECS (30 min) is based on the synaptic consolidation window described here.
The Neurobiology of Consolidations, or, How Stable Is the Engram?
2004
Dudai, Y.
Annual Review of Psychology, 55, 51-86
DOI: 10.1146/annurev.psych.55.090902.141555Annual Review of Psychology, 55, 51-86
How we use it: Detailed analysis of consolidation timelines. Informed our 24-hour session → long-term promotion threshold.
[07]
Graph-Enhanced Retrieval
Combining knowledge graphs with vector search
graph-enhanced-retrieval.viz
query → [semantic] ──┐
[keyword ] ──┤ RRF
[graph ] ──┤ fusion → results
[temporal] ──┘
multi-signal retrieval:
vector + BM25 + graph + recency
beats any single signal aloneFrom Local to Global: A Graph RAG Approach to Query-Focused Summarization
2024
Edge et al.
arXiv:2404.16130
DOI: 10.48550/arXiv.2404.16130arXiv:2404.16130
How we use it: GraphRAG improves retrieval by 13.1% over vector-only. Validated our hybrid semantic+graph approach.
A Syntactically-Based Query Reformulation Technique for Information Retrieval
2006
Lioma, C. & Ounis, I.
Information Processing & Management, 42(5), 1332-1363
DOI: 10.1016/j.ipm.2006.02.003Information Processing & Management, 42(5), 1332-1363
How we use it: Information content weighting for query terms. Our IC_NOUN, IC_ADJECTIVE, IC_VERB weights derive from this.
[08]
Associative Memory & Hopfield Networks
Content-addressable recall and energy-based memory retrieval—the foundation that transformers rediscovered
associative-memory-hopfield-networks.viz
Energy landscape:
╲ ╱╲ ╱╲ ╱
╲ ╱ ╲ ╱ ╲ ╱
╲ ╱ ╲ ╱ ╲ ╱
╲╱ ╲╱ ╲╱
mem_A mem_B mem_C
input rolls to nearest valley
partial query → complete recallNeural Networks and Physical Systems with Emergent Collective Computational Abilities
1982
Hopfield, J.J.
Proceedings of the National Academy of Sciences, 79(8), 2554-2558
DOI: 10.1073/pnas.79.8.2554Proceedings of the National Academy of Sciences, 79(8), 2554-2558
How we use it: The foundational paper on content-addressable memory. Shodh's core retrieval—query by similarity, converge to nearest stored memory—is a Hopfield recall operation over a knowledge graph.
Hopfield Networks is All You Need
2021
Ramsauer, H. et al.
International Conference on Learning Representations (ICLR)
DOI: 10.48550/arXiv.2008.02217International Conference on Learning Representations (ICLR)
How we use it: Proved transformer attention is mathematically equivalent to a modern Hopfield network update rule. Shodh's softmax-weighted retrieval scoring over stored memories operates on the same principle—energy minimization over an associative store.
Dense Associative Memories for Pattern Recognition
2016
Krotov, D. & Hopfield, J.J.
Advances in Neural Information Processing Systems (NeurIPS), 29
DOI: 10.48550/arXiv.1606.01164Advances in Neural Information Processing Systems (NeurIPS), 29
How we use it: Modern Hopfield networks with exponential storage capacity. Validates that associative memory scales—shodh's knowledge graph is an associative network with Hebbian-strengthened edges.
[09]
Biologically Plausible Learning
Local learning rules that don't need backpropagation—how brains actually learn
biologically-plausible-learning.viz
backpropagation: shodh/biological: global loss → chain local co-activation rule across all → strengthen edge layers (non-bio) (Hebbian, local) needs full graph needs only neighbors O(params) O(1) per update brain can't do this brain does exactly this
The Forward-Forward Algorithm: Some Preliminary Investigations
2022
Hinton, G.
arXiv:2212.13345
DOI: 10.48550/arXiv.2212.13345arXiv:2212.13345
How we use it: Hinton's alternative to backpropagation using local learning without backward passes. Shodh's Hebbian edge updates follow the same philosophy—purely local rules, no global error signal, biologically plausible weight adjustment.
Synaptic Plasticity Forms and Functions
2020
Magee, J.C. & Grienberger, C.
Annual Review of Neuroscience, 43, 95-117
DOI: 10.1146/annurev-neuro-090919-022842Annual Review of Neuroscience, 43, 95-117
How we use it: Comprehensive review of synaptic plasticity mechanisms including LTP phases. Directly implemented in shodh's 3-tier edge system: L1 Working (early-LTP, minutes), L2 Episodic (late-LTP, hours–days), L3 Semantic (structural LTP, permanent).
A History of Spike-Timing-Dependent Plasticity
2012
Markram, H. et al.
Frontiers in Synaptic Neuroscience, 3, 4
DOI: 10.3389/fnsyn.2011.00004Frontiers in Synaptic Neuroscience, 3, 4
How we use it: Documents the transition from rate-based to timing-based Hebbian learning. Shodh's co-retrieval strengthening (accessing memories in the same session strengthens their edge) implements the rate-based variant of STDP.
[10]
Vector Search & Approximate Nearest Neighbors
Efficient similarity search at scale without brute force
vector-search-approximate-nearest-neighbors.viz
Vamana graph (DiskANN):
●───●───● medoid
|\ | / | (entry point)
| ●─●─● | |
|/ | \ | v
●───●───● greedy walk to
nearest neighbor
<100k: Vamana | >100k: SPANN+PQDiskANN: Fast Accurate Billion-point Nearest Neighbor Search on a Single Node
2019
Subramanya, S.J. et al.
Advances in Neural Information Processing Systems (NeurIPS), 32
DOI: 10.48550/arXiv.1903.09556Advances in Neural Information Processing Systems (NeurIPS), 32
How we use it: The paper behind shodh's primary vector index. Our Vamana implementation uses DiskANN's greedy search with robust pruning for graph construction, enabling sub-millisecond ANN search on disk-resident data.
SPANN: Highly-efficient Billion-scale Approximate Nearest Neighborhood Search
2021
Chen, Q. et al.
Advances in Neural Information Processing Systems (NeurIPS), 34
DOI: 10.48550/arXiv.2111.08566Advances in Neural Information Processing Systems (NeurIPS), 34
How we use it: Shodh auto-switches from Vamana to SPANN when memory count exceeds 100k. SPANN's inverted index with product quantization enables billion-scale search with bounded memory via disk-based posting lists.
Product Quantization for Nearest Neighbor Search
2011
Jegou, H. et al.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(1), 117-128
DOI: 10.1109/TPAMI.2010.57IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(1), 117-128
How we use it: Foundation of the PQ compression in shodh's SPANN tier. 384-dim MiniLM vectors are decomposed into subspaces and quantized to codebook indices, reducing memory footprint by ~32x.
Open Science, Open Source
All 27 citations and 200+ tunable constants are documented in src/constants.rs with full justification. We believe AI memory systems should be grounded in science, not magic numbers.