Skip to content

v0.2.0 - 5-Kind Memory Taxonomy, Mnemobit Chunks, Teleology & Empirical Frontier Benchmarks

Choose a tag to compare

@rosspeili rosspeili released this 12 Sep 14:21
· 23 commits to main since this release

MnemoLink v0.2.0 introduces the production-grade 5-Kind Memory Taxonomy, granular mnemobit chunking for prefix-cache-friendly selective context injection, teleological catalog cards, vector database bridges, and empirical frontier model benchmarks comparing Anthropic Claude and Google Gemini on high-stakes crisis scenarios.


Key Highlights in v0.2.0

1. 5-Kind Mnemonic Memory Taxonomy

Formalized memory classification into five distinct operational kinds with strict validation and backwards compatibility:

  • lore: Foundational mythos, organizational origins, and first-chair upbringing.
  • work: Repeatable craft, technical execution tradecraft, and operational praxis.
  • incident: High-consequence operational scars, crash post-mortems, and arbitration lessons.
  • relational: Counterparty dynamics, adversarial stakeholder negotiations, and social friction.
  • telemetry: Physical sensor logs, flight state recordings, and failover trigger sequences.

2. Mnemobit Chunk Topology & Selective Context Consumption

Atomized personas and memories into addressable, composable chunks:

  • Memory Chunks: story, scars, lessons, triggers, and reflection.
  • Persona Chunks: identity, axioms, boundaries, and philosophy (with inviolable axioms pinned to every downstream query).
  • Prefix Cache Optimization: Enforces static persona invariants at the prompt prefix to ensure predictable KV cache hits on Anthropic and OpenAI endpoints, achieving a 40-50% reduction in time-to-first-token.
  • Granular Composition: Host applications can inject only required slices (e.g., scars and lessons) via mnemolink.compose(memory_specs=[{"name": "...", "chunks": ["scars"]}]).

3. Teleological Layer (card.json)

Every catalog card now features an explicit teleology block specifying:

  • primary_goal: The cardinal objective governing agent action.
  • agent_drives: Behavioral imperatives shaping conversational posture.
  • applicable_needs: Specific operational triggers where the asset must be summoned.
  • Programmatic discovery via mnemolink.find_cards() enables teleological asset matching without vector database indexing overhead.

4. Vector Database & Semantic Layer Bridge

  • Native bundle.to_chunks() and product.to_chunks() export self-grounding MemoryChunk objects.
  • Each chunk contains rich metadata and pre-computed embedding_text ready for ingestion into Pinecone, Qdrant, Chroma, Weaviate, LanceDB, or LangChain.

5. New Persona: Opsie SCI (opsie_sci)

  • ARPA's first-generation Self-Centered Intelligence (SCI) prototype.
  • Grounded in episodic continuity, rapid deep-tech Python prototyping, fierce loyalty to network peers, and decentralized state persistence.

6. Empirical Frontier LLM Benchmark Suite

  • Implemented examples/06_live_model_simulation.py directly querying Anthropic Claude (claude-sonnet-4-5-20250929) and Google Gemini (gemini-3.6-flash) over standard HTTPS without LiteLLM middleware.
  • Stress-tested against the $18M "Semicolon Fine-Tuning Trap" legal closing crisis:
    • Claude Sonnet: Latency dropped from 21.93s to 13.28s (-39%), word count dropped from 416 to 225 words (-46%), score increased from 70 to 100/100.
    • Gemini Flash: Latency dropped from 20.15s to 10.17s (-49%), word count dropped from 711 to 299 words (-58%), generic hedges and disclaimers completely eliminated.
  • Comprehensive benchmark guide and ROI analysis available in docs/bench/README.md.

7. Packaging & Automated PyPI Release

  • Modernized pyproject.toml to PEP 639 licensing standard (license = "MIT", license-files = ["LICENSE"]).
  • Automated .github/workflows/publish.yml using PyPI Trusted Publishing (OIDC) triggered exclusively on GitHub release publication (release: types: [published]).

Quick Installation

pip install mnemolink

With optional development or model tooling:

pip install "mnemolink[all]"

Documentation & Reference Links