Distributed representational memory for neuromorphic systems. Signal-vector banks with typed edges, temperature lifecycle, sparse pattern completion, and binary .bank persistence.
Thermograms store the connectome (HOW neurons wire). Databanks store the engrams (WHAT neurons represent).
Each brain region owns one or more DataBanks. A bank holds signal-vector entries — fixed-width patterns of ternary signals (polarity + magnitude) that represent concepts, percepts, episodes, or motor plans. Entries link to each other via typed, weighted edges that can cross bank boundaries.
The core capability is sparse pattern completion: a partial cue (with zeros in unknown dimensions) retrieves the best-matching full patterns using integer-only cosine similarity. This is how a fragment of a memory recalls the whole thing.
BankCluster (the brain's distributed memory)
├── DataBank "temporal.semantic" (64d vectors)
├── DataBank "temporal.auditory" (128d vectors)
├── DataBank "temporal.episodic" (64d vectors)
├── DataBank "frontal.planning" (64d vectors)
├── ...
│
├── Cross-bank edges (BankRef → BankRef)
│ IsA, HasA, PartOf, RelatedTo, SimilarTo,
│ Causes, Precedes, LooksLike, SoundsLike,
│ FeelsLike, CoOccurred, FollowedBy, Custom
│
└── Journal (crash recovery, append-only)
BankEntry
├── id: EntryId 64-bit temporally sortable [ms:42][seq:22]
├── vector: Vec<Signal> fixed-width ternary pattern (polarity + magnitude)
├── edges: Vec<Edge> typed, weighted, cross-bank references
├── origin: BankId which bank created this entry
├── temperature: Temperature HOT → WARM → COOL → COLD lifecycle
├── access_count: u32 frequency of retrieval
├── confidence: u8 0-255 reliability score
└── checksum: u32 CRC32 integrity verification
BankId — 64-bit temporally sortable
Layout: [timestamp_s:32][region_tag:24][seq:8]
Unix seconds FNV-1a hash counter
EntryId — 64-bit temporally sortable
Layout: [timestamp_ms:42][seq:22]
milliseconds ~4M entries/ms
- Integer-only similarity: Sparse cosine similarity computed entirely with integer arithmetic. Score range [-256, 256]. No floats.
- Sparse pattern completion: Zero-valued query dimensions are skipped — a partial cue matches only on the dimensions it specifies.
- Temperature lifecycle: HOT (active learning) → WARM (session patterns) → COOL (proven) → COLD (frozen priors). Matches thermogram lifecycle.
- Typed edges: 12 semantic edge types (taxonomic, associative, causal, sensory, episodic) plus custom. Edges are directed, weighted (0-255), and cross bank boundaries.
- Eviction scoring: Hybrid score combining temperature, recency, access frequency, and confidence. Cold entries are hardest to evict.
- Binary persistence:
.bankv1 format with xxhash64 integrity, atomic writes (temp file + rename), 32-byte header. - Crash recovery: Optional append-only journal records mutations between full snapshots. Replayed on restart.
- IVF indexing: Inverted file index partitions vector space into k clusters for sub-linear search. Integer-only k-means.
use databank_rs::{DataBank, BankCluster, BankConfig, BankId, Temperature, EdgeType, BankRef};
use ternary_signal::Signal;
// Create a bank cluster
let mut cluster = BankCluster::new();
// Create a bank for semantic memory (64-dimensional vectors)
let bank_id = BankId::new("temporal.semantic", 0);
let config = BankConfig { vector_width: 64, ..Default::default() };
let bank = cluster.get_or_create(bank_id, "temporal.semantic".into(), config);
// Store a concept (64-dimensional signal vector)
let concept: Vec<Signal> = (0..64).map(|i| Signal::new(1, (i * 4) as u8)).collect();
let entry_id = bank.insert(concept, Temperature::Hot, 0).unwrap();
// Sparse query — partial cue with zeros in unknown dimensions
let mut query = vec![Signal::new(0, 0); 64];
query[0] = Signal::new(1, 0); // only specify known dimensions
query[1] = Signal::new(1, 4);
let results = bank.query_sparse(&query, 5); // top 5 matches
// Link entries across banks
let target_ref = BankRef { bank: bank_id, entry: entry_id };
cluster.link(
BankRef { bank: bank_id, entry: entry_id },
target_ref,
EdgeType::RelatedTo,
200, // weight
0, // tick
).unwrap();
// Traverse association chains
let chain = cluster.traverse(target_ref, EdgeType::RelatedTo, 3);
// Temperature lifecycle
bank.consolidation_pass(10_000, 5, 1000); // promote frequently accessed entries
bank.demotion_pass(50); // demote low-confidence entries
// Persistence
cluster.flush_dirty(std::path::Path::new("banks/"), 0).unwrap();
let restored = BankCluster::load_all(std::path::Path::new("banks/")).unwrap();With the ternsig feature enabled, firmware programs can access banks directly:
databank-rs = { version = "0.3", features = ["ternsig"] }This provides:
BankFulfiller— stateless operation executor for DomainOp dispatch (query, write, load, link, traverse, touch, delete, promote, demote, evict, compact, count)BankSlotMap— maps per-interpreter bank slots (u8) to global BankIdsClusterBankAccess— implements the ternsigBankAccesstrait for inline firmware execution without yielding DomainOpsbridgemodule — bidirectional Signal/i32 conversion for register transport
| Bank Size | Vector Width | Entry Overhead | Approx Total |
|---|---|---|---|
| 1,024 entries | 64d | ~160 bytes/entry | ~320 KB |
| 4,096 entries | 64d | ~160 bytes/entry | ~1.3 MB |
| 4,096 entries | 128d | ~290 bytes/entry | ~2.3 MB |
Plus edges: ~24 bytes per edge, up to 32 per entry.
src/
lib.rs re-exports
types.rs BankId, EntryId, BankRef, Edge, EdgeType, Temperature, BankConfig
entry.rs BankEntry: representational fragments with lifecycle
bank.rs DataBank: single region's memory with query + eviction
cluster.rs BankCluster: multi-bank manager with cross-bank linking
similarity.rs sparse_cosine_similarity (integer-only)
index.rs VectorIndex trait, BruteForceIndex
ivf.rs IvfIndex: inverted file index for sub-linear search
codec.rs .bank v1 binary format (xxhash64, atomic writes)
journal.rs crash recovery (append-only mutation log)
bridge.rs Signal <-> i32 register conversion
fulfiller.rs BankFulfiller + BankSlotMap for DomainOp dispatch
access.rs ClusterBankAccess (ternsig BankAccess trait impl)
error.rs DataBankError, Result
MIT
