Mnemoverse Docs

Appearance

Glossary ​

This glossary defines canonical terms used across architecture, implementation, and documentation. Keep entries short and implementation‑agnostic.

Core Layers (L1–L8) ​

  • Knowledge Layer (L1) — aka Noosphere Layer: global knowledge store with quality/provenance and unified search abstraction (vector | graph | agents). Not project‑specific.
  • Project Library (L2): project‑scoped curated subset and references to L1; relevance‑first indexing and quick access (hot set). Different SLO from L1.
  • Workshop (L3): tool registry and validation pipelines (cold/stream); publishes signals; independent from concrete tool implementations.
  • Experience Layer (L4): task trails (task → nodes used → result → validation); searchable history of solved paths and outcomes.
  • Mnemoverse Memory (L5): on‑demand context assembly (budgets, LOD) with explicit source priorities and graceful degradation.
  • Agent Runtime (L6): agent planning/skills/execution (out of scope; integrated via adapters).
  • Interfaces & Clients (L7): IDEs/CLIs/assistants integrating via MCP/HTTP adapters.
  • Evaluation (L8): cross‑cutting benchmarks, metrics, and feedback observing L1–L7; offline‑first with controlled online actions.

Data & Metadata ​

  • Document: any retrievable unit (paper, spec, code file, note) with id, title, uri, summary.
  • Provenance: origin metadata (source, author, acquisition method, timestamps).
  • Quality Signals: trust/reliability indicators (source reputation, citation strength, moderation flags).
  • Reference (L2): link from project library to an L1 document with project‑owned fields (relevance, notes, freshness state).
  • Freshness: staleness/updatedness status; may be time‑ or event‑driven.
  • Hot Set: subset optimized for low‑latency access in L2 (recent, frequent, critical).
  • Working Set: task‑scoped selection of items in active use.

Retrieval, Ranking, Routing ​

  • Unified Search Abstraction: single API that routes to vector/graph/agent methods and returns normalized results + routing metadata.
  • Vector Search: embedding‑based semantic retrieval; chunking + embedding + ANN index.
  • Graph Search: traversal/paths over knowledge graph; relations, centrality, communities.
  • Reranking: secondary scoring stage (ML/heuristics) to refine initial results.
  • Router: logic selecting method(s) per request; may consult cache and feedback.
  • Pattern Cache: cache keyed by request features → method sequence/parameters + success stats (LFU+recency, TTL).

Context Assembly (L5) ​

  • Budget Policy: constraints for tokens/time/IO in assembling context; hard/soft limits.
  • LOD (Level of Detail): granularity level (coarse → fine) for included items/snippets/modalities.
  • Source Priorities: default order L2 > L4 hints > L1; configurable per environment.
  • Degradation: deterministic fallback when budgets/timeouts hit; partial context with provenance.

Feedback, Telemetry, Privacy ​

  • Feedback Signals: utility indicators (like/dislike, utility score, dwell time, clicks) without logging sensitive payloads.
  • Telemetry Modes: Off (no store), Minimal (aggregates, pseudonymous IDs), Enhanced (opt‑in extended metrics).
  • Pseudonymization: hashing/pseudonyms for user/project IDs; keys remain with org/user.
  • Retention: raw events kept 30–90 days; aggregates longer; deletion supported.

Interfaces & Contracts ​

  • MCP Adapter: Model Context Protocol commands/payloads bridging clients and layers.
  • HTTP Adapter: REST endpoints for programmatic access.
  • Contract: documented request/response schemas, error codes, and SLO expectations between layers.
  • Policy: normative rules (priorities, budgets, freshness) applied by components.

Orchestration Components (Cross‑Cutting) ​

  • ACS (Adaptive Context Scaling): planner that selects “what and at which LOD” under budgets; outputs fragments and KV policy; reusable across retrieval surfaces (not a single layer).
  • CEO (Context/Execution Orchestrator): origin of intent, budgets, and risk profile; issues render requests to ACS; collects metrics back from HCS.
  • HCS (Human‑Centric Streaming): streaming/transport layer (SSE/WS) delivering fragments progressively with pacing.
  • Pacing: progressive|burst|drip|guarded delivery strategies chosen per request.
  • KV Policy: client cache guidance (pin/compress/evict) for keys/snippets to control reuse and memory pressure.
  • Portals: navigational links between fragments (uses/explains/evidence) to enable focused exploration.

Metrics & Evaluation ​

  • Latency p50/p95/p99: response time percentiles per method/layer.
  • Precision@K / Recall / NDCG@K: retrieval quality metrics.
  • Cache Hit Rate: fraction of requests served from pattern cache.
  • Utility Score: normalized measure (0–1) summarizing feedback signals.

Error Handling ​

  • Timeout: exceeded time budget; triggers degradation/fallback.
  • Retry & Backoff: controlled re‑attempt policies; avoid cascading failures.
  • Circuit Breaker: prevent repeated calls to failing component.

Naming & Conventions ​

  • English terms in docs and APIs; consistent layer labels (L1–L8).
  • Use kebab‑case for filenames, snake_case for JSON fields (unless client requires otherwise).
  • All IDs are stable and opaque (no meaning encoded).