Kernel Shapes

🔷 Typed Memory Shapes

Uniform physical access abstractions separating structural access patterns from concrete memory store semantics.

---

The Shape Principle

In low-level storage architectures, coupling the physical storage layout with domain logic creates brittle systems. For example, treating "Semantic Memory" as a completely unique data structure obscures the fact that at the byte level, it shares identical access patterns with "Working Memory" and "Procedural Memory".

The Spector Memory Kernel resolves this via The Shape Principle:

Memory Shapes define access patterns, not storage semantics. Concrete cognitive subsystems name their domain content only (SemanticMemory, HebbianGraph, StrengthLayout), while the kernel executes operations through ten canonical Memory Shapes.

graph TD
    subgraph "Canonical Memory Shapes"
        RM["RecordMemory<br/><i>Contiguous Indexed Slots</i>"]
        AM["AppendMemory<br/><i>Sequential Append Stream</i>"]
        GM["GraphMemory<br/><i>CSR Adjacency Matrix</i>"]
        CM["ChainMemory<br/><i>Chronological Linked Nodes</i>"]
        HM["HashTableMemory<br/><i>O(1) Direct Key-Value Map</i>"]
        YM["RegistryMemory<br/><i>Bidirectional Symbol Interning</i>"]
        EM["EntityDirectoryMemory<br/><i>Entity & Role Identification</i>"]
        IM["InsulaMemory<br/><i>Somatic Self-Model Container</i>"]
    end

    subgraph "Cognitive Subsystem Backing"
        T1["Working, Semantic, & Procedural Tiers<br/><i>Fixed-stride engram slots</i>"]
        T2["Write-Ahead Log (WAL) & Text Blobs<br/><i>Append-only cursor log</i>"]
        T3["Hebbian Graph & HyperEntity Graph<br/><i>Synaptic associative edges</i>"]
        T4["Temporal Episode Sequence<br/><i>Chronological narrative links</i>"]
        T5["Pairwise Co-Activation Tracker<br/><i>Fast frequency lookups</i>"]
        T6["Entity & Relation Type Namespaces<br/><i>Dynamic schema interning</i>"]
        T7["Agent Identity & External Entities<br/><i>Entity directories</i>"]
        T8["Dynamic Self-Model &amp; Self-State<br/><i>Single-entry JSON self-state</i>"]
    end

    RM --> T1
    AM --> T2
    GM --> T3
    CM --> T4
    HM --> T5
    YM --> T6
    EM --> T7
    IM --> T8

Loading

---

Detailed Shape Specifications

1. RecordMemory (Contiguous Indexed Slots)

RecordMemory manages contiguous, fixed-stride off-heap byte slots indexed by sequential numerical identifiers.

• Access Pattern: Random read/write access via direct slot offset calculations: $\text{Offset} = \text{Base} + (\text{Index} \times \text{Stride})$.
• Concurrency: Lock-free concurrent reads across Virtual Threads using hardware atomic primitives (VarHandle compare-and-swap and volatile loads).
• Backed Subsystems:
  • WorkingMemory (circular buffer)
  • SemanticMemory (crystallized concepts)
  • ProceduralMemory (learned operational skills)
  • StrengthMemory (96-byte recall dynamics and Bjork storage strength)

---

2. AppendMemory (Sequential Append Stream)

AppendMemory provides an append-only stream of variable- or fixed-length records controlled by an atomic cursor pointer.

• Access Pattern: Unidirectional sequential appends with monotonic sequence counters.
• Concurrency: Single-writer or lock-coordinated sequential writes; concurrent non-blocking sequential reads.
• Backed Subsystems:
  • WalMemory (Write-Ahead Log event records)
  • TextBlobMemory (raw document text payloads)
  • TemporalFactMemory (bi-temporal assertion streams)

---

3. GraphMemory (Adjacency Matrices & Slabs)

GraphMemory implements a high-performance off-heap graph engine using Compressed Sparse Row (CSR) layouts combined with dynamic adjacency slabs.

• Access Pattern: Traversal of outgoing and incoming edges, weight adjustments, and neighbor queries without creating heap objects.
• Concurrency: Read-mostly graph traversals optimized for SIMD operations, with thread-safe atomic edge weight updates.
• Backed Subsystems:
  • HebbianMemory (associative synaptic connections between memory engrams)
  • HyperGraphMemory ($n$-ary hyperedges linking entities, roles, and contexts)

---

4. ChainMemory (Chronological Linking)

ChainMemory maintains ordered sequences of temporal events linking conversations, episodes, and cognitive checkpoints.

• Access Pattern: Forward and backward traversals along causal chains: "what happened immediately before?" and "what occurred next?".
• Characteristics: Compact pointer offsets linking prior and subsequent engram indices across session boundaries.
• Backed Subsystems:
  • TemporalChainMemory (episodic narrative sequencing)

---

5. HashTableMemory (Off-Heap Key-Value Map)

HashTableMemory is a low-level, collision-resilient hash table residing entirely in off-heap memory.

• Access Pattern: $O(1)$ key lookup and increment operations using open addressing with linear probing.
• Characteristics: Eliminates Java heap Map overhead (which typically incurs 32–48 bytes per entry in object headers and references).
• Backed Subsystems:
  • CoActivationMemory (tracks pairwise co-retrieval frequencies between engrams to dynamically reinforce associative bonds)

---

6. RegistryMemory (Symbol Interning)

RegistryMemory establishes a two-way mapping between variable-length string symbols and compact integer identifiers.

• Access Pattern: Fast bidirectional resolution: String $\rightarrow$ Integer ID and Integer ID $\rightarrow$ String.
• Characteristics: Enables open-schema architectures where new entity types, relation names, and tags are registered dynamically without storing repetitive string tokens across the hot off-heap memory path.
• Backed Subsystems:
  • EntityTypeMemory (interned entity categories: person, organization, location, etc.)
  • RelationTypeMemory (interned relationship predicates: causes, depends_on, works_at, etc.)

---

7. EntityDirectoryMemory (Entity Directory)

EntityDirectoryMemory coordinates global entity identifiers, alias tracking, and name pool mappings across a namespace.

• Access Pattern: Fast entity lookup, type resolution, and association mapping.
• Backed Subsystems:
  • Inter-agent entity resolution, cross-session participant identification, and structured hypergraph entity registries.

---

8. InsulaMemory (self-model region used by AISME — experimental) (Dynamic Self-Model Container)

InsulaMemory implements a dedicated single-entry container storing the agent's dynamic, real-time self-model and runtime state.

• Access Pattern: Single-entry atomic read/write (put(), get(), clear()) of a variable-length JSON self-model payload.
• Design Rationale: A single, versioned JSON document holds the agent's confidence, uncertainty, and affective state — updated in sub-microsecond cycles during active reasoning, without scanning or indexing.
• Integrity & Concurrency:
  • Sub-header tracks a monotonic version counter, payload byte length, epoch timestamp, and a hardware-computed CRC-32C checksum.
  • Concurrency is protected by a thread-safe write lock with zero-copy unaligned read semantics across concurrent Virtual Threads.
• Backed Subsystems:
  • InsulaMemory (active self-model, task confidence, dynamic stress level, and salience posture within runtime.bundle at RegionId.INSULA).

---

Shape Comparison Summary

Memory Shape	Record Length	Indexing	Primary Operations	Off-Heap Layout
RecordMemory	Fixed stride	Numeric slot index	write(long, MemorySegment), read(long, MemorySegment), recordOffset(long)	Contiguous cache-aligned array
AppendMemory	Variable / Fixed	Sequential byte offset	append(MemorySegment), read(long, int), replay(long), appendCursor()	Cursor-driven append log
GraphMemory	Fixed node/edge	Node identifier	addEdge(int, int, MemorySegment), removeEdge(int), neighbours(int), edgeCount(), nodeCount()	Compressed Sparse Row + slab lists
ChainMemory	Fixed node	Link identifier	link(int, int), next(int), prev(int), head(), tail(), chainLength()	Bidirectional offset pointers
HashTableMemory	Fixed entry	Hash key	Marker interface — concrete operations defined by CoActivationMemory	Open-addressing probing buffer
RegistryMemory	Variable	String hash & ID	intern(String), putDirect(String, int), nameOf(int), idOf(String), entries()	String pool + ID lookup table
EntityDirectoryMemory	Composite	Entity identifier	directoryRegion(), namesRegion(), entityCapacity(), entityCount(), flush(), close()	Directory index + name buffer
InsulaMemory	Variable (1 slot)	Singleton self-model	put(byte[]), get(), clear(), version(), isPresent()	32-byte header + CRC-32C JSON payload
BundleMemory	Multi-region	Region directory	Container lifecycle (open, map, close)	Multi-region mmap container (ADR-0004)
DistributedMemoryTensor	Fixed tensor	Holographic	accumulate(float[], float), retract(float[], float), evaluateEnergy(float[], float), decay(float)	64-byte header + off-heap PRF accumulator (ADR-0020)

Note

PARTITIONED is retained in the MemoryShape enum for on-disk compatibility (ordinal stability) but has no live code references — the PartitionedRecordMemory abstraction was removed as dead code.