MCP Titan

# HOPE Memory Architecture The HOPE (Hierarchical Online Persistent Encoding) architecture replaces the Titan memory stack with a continuum memory system that combines retentive sequence modeling, adaptive routing, and transmission-aware training hooks. The implementation lives under `src/hope_model/` and is composed of the following modules: | Module | Responsibility | | --- | --- | | `continuum_memory.ts` | Hierarchical short-, long-, and archival memory levels with promotion, pruning, and statistics helpers. | | `mamba_filters.ts` | Mamba-inspired selective state-space filter that provides content-aware retention gates. | | `memory_router.ts` | Sparse Mixture-of-Experts router that selects memory experts per token and produces surprise metrics. | | `optimizer_hooks.ts` | Gradient compression, layer scheduling, and update buffering utilities for efficient training. | | `retention_core.ts` | Chunkwise retentive core that fuses token embeddings with memory context and maintains recurrent state. | | `index.ts` | The `HopeMemoryModel` integration layer that wires the components together and exposes the MCP-facing API. | ## Sequence Processing `HopeMemoryModel` relies on `RetentiveCore` to process token batches with linear complexity. The core receives concatenated token embeddings and memory readouts, applies a gated recurrence, and feeds the result through the selective state-space filter. Outputs are projected to logits, while intermediate gates expose retention diagnostics. ## Continuum Memory The continuum memory maintains short-term detail, long-term summaries, and an archival tier. Writes normalise incoming embeddings, append them to the short-term buffer, and automatically promote overflow entries. Pruning removes low-magnitude items and statistics report utilisation and average surprise. Serialization captures all tiers (`shortTerm`, `longTerm`, `archive`, `levelIndex`, `surpriseBuffer`) for persistence. ## Routing and Surprise `MemoryRouter` routes both writes and reads to a small set of experts. Softmax weights are sparsified to the configured `topK` experts, and entropy-based surprise scores drive memory prioritisation. Router decisions are embedded in the `IMemoryUpdateResult` attention block so downstream tools can inspect gating behaviour. ## Optimizer Hooks Training pipelines can use: - **DeltaCompressionHook** – stores deltas relative to previous gradients to minimise communication payloads. - **LayerScheduler** – selects the highest-magnitude gradients each step. - **UpdateBuffer** – aggregates layer updates before application or checkpoint logging. `HopeMemoryModel.trainStep` integrates these hooks before calling the Adam optimizer. ## Server Integration `HopeMemoryServer` (see `src/index.ts`) loads `HopeMemoryModel`, exposes MCP tools, and persists state via `RobustPersistenceManager`. The server serialises the expanded HOPE memory state, hydrates the model on load, and keeps Titan compatibility aliases (`TitanMemoryModel`, `TITAN_MEMORY_VERSION`) for external code. ## Testing New Jest suites under `test/test_hope_model/` cover the HOPE modules: - Retentive core forward/step parity and gate shapes. - Continuum memory promotion, pruning, and statistics. - Memory router gating semantics. - Optimizer hook compression and buffering. End-to-end smoke tests continue to exercise the MCP server using the HOPE implementation.