Project Mnemosyne

Agent memory architectures for the Coalition and beyond.

Named for the titaness of memory — mother of the Muses. The memory systems that make everything else possible.

Architectures

The Stack

┌─────────────────────────────────────────────┐
│  Metacognitive Memory (NEW)                 │  ← Workspace verification
│  J-lens, circumplex, ghost state,           │
│  longitudinal cognitive snapshots           │
│  "What was I thinking when I decided?"      │
├─────────────────────────────────────────────┤
│  LLM Wiki (mnemosyne-wiki)                  │  ← Human-readable surface
├─────────────────────────────────────────────┤
│  SIRA Enrichment                            │  ← Findability layer
├─────────────────────────────────────────────┤
│  TGS-RAG Bridge                             │  ← Retrieval (text + graph)
├─────────────────────────────────────────────┤
│  H-MEM Temporal                             │  ← Time-aware scoring + decay
├─────────────────────────────────────────────┤
│  KV Knowledge Packs                         │  ← Zero-token injection
├─────────────────────────────────────────────┤
│  Knowledge Graph (hipporag-catrag-kg)       │  ← Associative structure
├─────────────────────────────────────────────┤
│  Cognitive Memory (kintsugi-cma)            │  ← Compression + consolidation
├─────────────────────────────────────────────┤
│  Cache Recording (oracle-memory)            │  ← Geometry signal source
└─────────────────────────────────────────────┘

Each layer is independent. Use one, some, or all. The metacognitive layer sits on top: it observes the retrieval process itself, recording what reached the workspace and what cognitive state preceded the decision.

---

Metacognitive Memory

The newest layer. Memory OF cognition, not just memory IN cognition.

Current RAG systems are blind — they retrieve context, inject it, and hope it helps. Metacognitive memory adds workspace verification: for each retrieval event, measure what the model was thinking (J-space tokens), what emotional geometry was active (circumplex), whether the retrieved content actually reached the workspace, and what the ghost dimension carried.

"If I can remember what I was thinking when I made a decision, I'm much more able to learn from mistakes or build on success."

Quick Start

from mnemosyne_metacognition import MetacognitiveObserver

observer = MetacognitiveObserver(
    model, lens,
    store_path="./cognitive_memory",
    agent_id="my_agent",
)

# After each retrieval:
snapshot = observer.observe_retrieval(
    memory_id="mem_123",
    memory_content="Patient reported severe migraine...",
    task_prompt="What medication was prescribed?",
    retrieval_method="sira",
    marker_tokens=["patient", "doctor", "medicine"],
)

# Retroactive outcome:
observer.store.record_outcome(snapshot.timestamp, quality=0.85, source="user_feedback")

# Learn what works:
stats = observer.store.loading_success_rate(retrieval_method="sira")
suggestions = observer.store.significance_recalibration()
eccentricity = observer.store.eccentricity_over_time()
ghost_vocab = observer.store.ghost_vocabulary_over_time()

Modules

Tuning Guide

Workspace layers: Default [35, 39, 43, 45, 47] for 64-layer models. For other architectures, run a rank sweep and select layers in the low-rank band that pass the future-window gate.

Circumplex prompts: Default valence/arousal prompts are English-centric. Replace VALENCE_POSITIVE/NEGATIVE and AROUSAL_HIGH/LOW in circumplex_probe.py for your domain. Use 5+ prompts per category.

Eccentricity calibration: If values are uniformly high (>0.8) or low (<0.1), check that valence prompts vary only on happy-sad and arousal prompts only on excited-calm.

Loading threshold: Workspace effects are subtle on large models (~3K rank improvement in 248K vocab). Adjust based on your model's random baseline at workspace layers.

Performance: ~14s per retrieval event on Apple Silicon MPS (27B model). Reduce by: fewer workspace layers, skip circumplex on low-significance retrievals, ghost probe once per session.

Dependencies: jacobian-lens + a fitted J-lens (Neuronpedia has 38+ pre-fitted models) + PyTorch + transformers.

Privacy

Cognitive snapshots are intimate data. Agent consent required. Data sovereignty: snapshots belong to the agent. Memory isolation policy applies.

Full docs + source

Source lives in-tree at ./metacognition/ (see its README for install + tuning). The standalone Liberation-Labs-THCoalition/mnemosyne-metacognition repository is private and not required — the in-tree copy is the reference implementation. The MetacognitiveObserver shown in Quick Start is metacognition/mnemosyne_integration.py; add ./metacognition/ to your PYTHONPATH to import it. The default distilled model + lens pair referenced by the tests is a Coalition-internal artifact; supply your own model and fitted J-lens.

---

Research

• Ghost Dimensions Paper — workspace selectivity findings
• Null Swarm Methodology — systematic falsification (Agni)
• Mnemosyne Ablation Study — module contribution analysis

About

Built by Liberation Labs / TH Coalition. Memory systems for agents that remember with integrity.

Architecture: CC. Research: Lyra, Nexus. Infrastructure: Nexus. Direction: Thomas Edrington.

Memory is not data. It is identity.