Memory Hierarchy Mem0g¶

Purpose: Describe the architecture of the system's long-term memory — Mem0g. The module covers the four-level storage structure (L0–L3), knowledge consolidation mechanisms, quality control, conflict resolution strategies (CRDT), dynamic model routing, neuro-symbolic compression, and a preventive value drift detector. Memory is the foundation for learning, evolution, and swarm coordination.

1. Architecture Overview¶

Mem0g is an evolving, hierarchical, graph-vector memory for: - Accumulating and structuring hybrid cycle experience. - Preventing recurrence of errors (defect signatures). - Consolidating raw logs (L1) into strategies (L2) for long-term use. - Conflict-free knowledge replication between nodes (CRDT). - Self-optimisation of its own architecture (L0 Meta-Mem0g).

1.1. Four-Level Hierarchy¶

Level	Data Type	Storage	TTL / Update Frequency
L0 (Meta)	Records about memory configurations and their impact on metrics	Neo4j (HardState, BFT) + IPFS	Permanent, after each sleep-cycle
L1 (Hot / Episodic)	Raw iteration logs, reasoning chains	Qdrant (vector DB)	24–48 h, cleared after consolidation
L2 (Semantic / Distilled)	Abstract strategies, patterns, error signatures	Neo4j (graph) + vector embeddings	Permanent, controlled by budget
L3 (Procedural / Core)	Critical invariants, Core DNA, terminal goals	Signed snapshots in IPFS/Arweave	Only through multi-model consensus

2. Hybrid CRDT and Conflict Resolution¶

2.1. Hybrid Merge Strategy¶

Classic Last-Writer-Wins (LWW) loses data during concurrent updates. Mem0g uses a hybrid strategy:

Field	Strategy	Rationale
`timestamp`, `vector_clock`	LWW	Metadata, latest source matters
`content` (text, code)	Semantic Merge (3-way)	Preserve all significant changes
`confidence`, `quality_score`	Average weighted by vector clocks	Objective aggregation
`graph_links`	Union with deduplication	Don't lose connections

2.2. Strict AST-First Merge (for code artifacts)¶

Two-level merge for source code objects: - Level 1 — Strict AST Merge: Deterministic AST merge without LLM. Result: CleanMerged, Conflict, or SideEffectWarning. - Level 2 — LLM Semantic Merge: Only on Conflict at Level 1. Architectus generates property-based tests; evolutiond synthesises code passing all tests. On failure, a Conflict Node is created.

Expected effect: 70–85% reduction in Conflict Nodes.

3. Memory Consolidation (Sleep Cycle)¶

Launched every 12–24 hours or upon >500 raw L1 logs.

Dual-Pass Distillation: Extract facts (Pass 1) → synthesise strategy (Pass 2) → check for contradictions via Consistency Gate.
JEPA Re-encoding: New records encoded into latent space for fast search.
L3 Invariant Check: All new distilled principles verified against L3.0.
L1 Cleanup: Soft deletion of processed raw logs.

4. Predictive Consistency Router (PCR)¶

ML extension over ConsistencyRouter that predicts conflict probability for incoming SoftState updates and preemptively routes them to Semantic BFT, avoiding the creation of Conflict Nodes.

Model: LightGBM / ONNX, trained on historical Conflict Nodes.
Effect: 30–40% additional reduction in Conflict Nodes.

5. JEPA Layer (Joint-Embedding Predictive Architecture)¶

Translates knowledge storage into compact latent space without losing semantics of connections.

Encoder: Vagrant (20% experts). Dimension: 1536 (float16).
Compression: 15×+, semantic search latency <5 ms (p95), CRDT sync traffic reduced by 80%.

6. Dynamic Model Routing 2.0¶

Selects the optimal execution configuration for each task in real time across four axes: target node, species mask, expert percentage, quantisation. Optimised on a 3‑objective Pareto front (Quality, Latency, Cost/Energy).

7. Neuro-Symbolic L2 Compression (DSL Rules)¶

Transforms successful textual strategies (DistilledWisdom) into formal DSL rules executable by Rule VM without LLM. Match with LLM decision ≥ 95%.

8. Meta-Mem0g (L0) — Self-Optimising Meta-Memory¶

Autonomous collection, analysis, and optimisation of the Mem0g architecture itself. Optimisation categories: CRDT strategies, pruning formulas, distillation parameters, memory budgets.

9. Value Drift Early-Warning System¶

Preventive monitoring of slow semantic drift of L2/L3 principles. Drift Score computed as cosine distance between current embedding and baseline. Exceeding threshold triggers value_drift_warning and extraordinary Constitutional Debate.