Skip to content

ADR-0008: Cognitive Substrate Evolution (TANGLE, GPM, MSCE)

Field Value
Status Accepted (Implemented)
Date 2026-08-08
Authors Spector Maintainers & Architecture Working Group
Deciders Spector Technical Steering Committee (TSC)
Supersedes None
Superseded By None
Last Verified 2026-09-16 (Verified against main)

1. Context

As Spector evolved from a vector memory store into a comprehensive cognitive architecture, agent applications required richer cognitive substrates: temporal sequence tracking, procedural habit execution, homeostatic balance, and dynamic prompt fusion. To support autonomous persona replication and digital continuity, Spector established three specialized memory components: TANGLE (Temporal Associative Network Graph), GPM (Goal & Procedural Memory), and MSCE (Multi-Stage Context Extraction).

2. Problem Statement

Previous cognitive retrieval presented critical structural gaps:

  1. Static prompt formatting: Client applications had to manually fetch and stitch working memory, semantic facts, and episodic context, resulting in fragmented, redundant LLM prompts.
  2. Procedural amnesia: The system had no representation for behavioral heuristics, interaction cadences, or tool invocation policies.
  3. Temporal blind spots: Pure semantic search retrieved disjoint facts without awareness of chronological order or causal transitions.

3. Decision Drivers

  • Unified Context Synthesis: Provide a single atomic operation to retrieve fully formatted, multi-tier context packs directly consumable by LLMs.
  • Procedural Habit Representation: Model procedural memory (rules, heuristics, cadences) as active cognitive policies.
  • Temporal Causal Graphing: Maintain chronological episodic sequences and causal transition links.
  • Sub-100ms Latency Budget: End-to-end multi-tier context assembly must complete within strict real-time conversational budgets.

4. Considered Options

Option 1: Client-Side Multi-Query Fusion

  • Description: Clients make separate calls to episodic, semantic, and graph endpoints and merge prompts in client SDKs.
  • Advantages: Keeps server-side APIs minimal.
  • Disadvantages: Incurs multiple network roundtrips (150–400ms); client-side prompt formatting varies across languages and lacks unified scoring.

Option 2: Standard Graph-RAG Pipeline

  • Description: Use off-the-shelf Graph-RAG frameworks (NetworkX, external vector DBs).
  • Advantages: Utilizes existing open-source libraries.
  • Disadvantages: High GC pressure, high query latency (50–500ms), and inability to enforce biological cognitive constraints.

Option 3: Unified Cognitive Substrate Architecture (TANGLE, GPM, MSCE) (Selected)

  • Description: Implement an off-heap cognitive substrate natively inside spector-memory and spector-synapse:
    • TANGLE: Off-heap temporal associative network tracking chronological chains and entity co-activations.
    • GPM: Procedural heuristics engine representing interaction cadence and behavioral rules.
    • MSCE: Multi-Stage Context Extraction engine fusing working scratchpads, procedural habits, semantic axioms, and episodic anecdotes into a single coherent prompt payload.
  • Advantages: Delivers end-to-end prompt synthesis in < 15ms; guarantees consistent cognitive formatting across all agent runtimes.
  • Disadvantages: Requires deep integration across all memory tiers and synapse gateway layers.

5. Decision Outcome

Chosen Option: Option 3 (Unified Cognitive Substrate Architecture).

Positive Consequences

  • Single atomic API call generates complete, deduplicated cognitive prompt contexts for LLM agents.
  • Procedural memory enables agents to preserve unique behavioral reflexes and interaction cadences over time.
  • TANGLE graph ensures temporal and causal continuity across multi-session interactions.

Negative Consequences & Trade-offs

  • Increased complexity in synapse gateway routing to coordinate multi-tier context extraction.
  • Context pack generation requires strict token budgeting to prevent prompt overflow.

6. Pros and Cons of the Options

Option Pros Cons
Option 1: Client Fusion Simple server APIs High network latency, fragmented prompt structure
Option 2: External Graph-RAG Off-the-shelf code Heavy GC pauses, high query latency, no procedural tier
Option 3: Unified Substrate < 15ms latency, atomic context packs, procedural memory Deeper internal integration complexity

7. Implementation Plan

  1. Phase 1: Implement TangleGraphEngine for temporal episodic chaining in memory/spector-memory.
  2. Phase 2: Construct ProceduralMemoryStore for behavioral heuristics and interaction policies.
  3. Phase 3: Build MultiStageContextExtractor in synapse/spector-synapse to generate unified context packs.
  4. Phase 4: Wire context packs into MCP server tools and client SDK interfaces.

8. Code Reference & Verification

  • Primary Module(s): memory/spector-memory, synapse/spector-synapse
  • Key Packages: com.spectrayan.spector.memory.tangle, com.spectrayan.spector.synapse.context
  • Classes: TangleGraphEngine.java, ProceduralMemoryStore.java, MultiStageContextExtractor.java, ContextPack.java
  • Verification Tests: TangleGraphIntegrationTest.java, ContextExtractionPipelineTest.java