V3.2 Overview

V3.2 Overview: Associative Memory Upgrade

SuperLocalMemory v3.2 (codename "The Living Brain") introduces four capabilities that transform SLM from a retrieval engine into an associative memory system. Memories now surface automatically, form connections, detect contradictions, and consolidate during idle time --- all without requiring cloud LLM calls in Mode A.

These features are grounded in published research: the SYNAPSE spreading activation algorithm (arXiv 2601.02744), the A-MEM auto-linking architecture (agiresearch/A-mem-sys), ACT-R cognitive activation theory (Anderson & Lebiere 1998), and Letta's sleep-time consolidation model. For the mathematical foundations of the existing retrieval system, see the V3 paper.

What's New

Feature	What It Does	Key Innovation
Auto-Invoke	Memories surface automatically using 4-signal scoring	FOK gating rejects noise; optional ACT-R cognitive model
Association Graph	Multi-hop reasoning via spreading activation	SYNAPSE 5-step algorithm as a 5th retrieval channel
Temporal Intelligence	Bi-temporal validity and contradiction detection	4-timestamp model; sheaf cohomology for contradiction detection
Consolidation	Sleep-time memory consolidation with Core Memory blocks	Always-in-context working memory; rules-based compilation in Mode A

Feature Matrix by Mode

Every v3.2 feature respects the existing three-mode architecture. Mode A remains fully local and EU AI Act compliant.

Feature	Mode A (Local Guardian)	Mode B (Smart Local)	Mode C (Full Power)
Auto-Invoke scoring	3-signal (recency + frequency + trust; similarity=0 without embeddings)	4-signal (all signals active)	4-signal (all signals active)
Contextual descriptions	Template-based rules	Ollama LLM generation	Cloud LLM generation
Spreading activation	Disabled without embeddings; enabled with sentence-transformers	Full SYNAPSE algorithm	Full SYNAPSE algorithm
Auto-linking	Disabled without embeddings	Cosine similarity >= 0.7	Cosine similarity >= 0.7
PageRank + communities	Enabled (pure graph math)	Enabled	Enabled
Contradiction detection	Sheaf cohomology (pure linear algebra)	LLM verification with sheaf pre-filter	LLM verification with sheaf pre-filter
Temporal validity	Full bi-temporal (4 timestamps)	Full bi-temporal	Full bi-temporal
Consolidation	Rules-based block compilation (no LLM)	Ollama summarization	Cloud LLM summarization
Core Memory blocks	Top-N facts by access count and trust	LLM-assisted summaries	LLM-assisted summaries

Migration: Zero Breaking Changes

V3.2 is a strict superset of v3.1. Every feature is opt-in via configuration flags that default to false.

• All 29 MCP tools retain their existing signatures
• All 21 CLI commands work identically
• All 65+ API endpoints are unchanged
• Existing retrieval behavior is preserved when features are disabled
• 9 new database tables are created additively (no schema alterations)
• Rollback SQL is provided for every new table

To enable v3.2 features after upgrading:

slm config set auto_invoke.enabled true
slm config set spreading_activation.enabled true
slm config set temporal_validator.enabled true
slm config set consolidation.enabled true

Or enable them individually as you evaluate each capability.

Architecture

V3.2 adds 4 new modules that integrate with the existing engine via dependency injection. No circular imports. The engine remains a thin facade.

                        +-------------------+
                        |   Engine Facade   |
                        +-------------------+
                               |
         +----------+----------+----------+----------+
         |          |          |          |          |
    AutoInvoker  Spreading  AutoLinker  Temporal  Consolidation
    (Phase 2)   Activation  (Phase 3)  Validator   Engine
                (Phase 3)              (Phase 4)   (Phase 5)
         |          |          |          |
         +----+-----+----------+          |
              |                           |
        ChannelRegistry            fact_temporal_validity
        (5th channel)              (retrieval filter)

New database tables (9 total):

Table	Phase	Purpose
fact_access_log	1	Tracks when facts are accessed (recency scoring)
fact_embeddings (vec0)	1	Native vector search via sqlite-vec
embedding_metadata	1	Embedding model and dimension per fact
fact_context	2	Contextual descriptions (WHY a memory matters)
association_edges	3	Auto-linking and Hebbian co-access edges
activation_cache	3	Transient spreading activation results (1-hour TTL)
fact_importance	3	PageRank scores and community IDs
fact_temporal_validity	4	Bi-temporal validity (4 timestamps)
core_memory_blocks	5	Always-in-context working memory

Research References

• SLM V3 Paper: Information-Geometric Foundations for Agent Memory
• SYNAPSE: Spreading activation for episodic-semantic memory (arXiv 2601.02744)
• A-MEM: Agentic memory with auto-linking (agiresearch/A-mem-sys)
• ACT-R: Adaptive Control of Thought -- Rational (Anderson & Lebiere 1998)
• Letta: Sleep-time consolidation and Core Memory blocks
• Zep/Graphiti: Bi-temporal knowledge graphs for agent memory

Feature Pages

• Auto-Invoke --- Multi-signal automatic memory surfacing
• Association Graph --- Spreading activation and auto-linking
• Temporal Intelligence --- Bi-temporal validity and contradiction detection
• Consolidation --- Sleep-time memory consolidation

---

Part of Qualixar | Created by Varun Pratap Bhardwaj