MnemoCore Roadmap
Scope and Intent
This roadmap describes current known gaps and likely direction. It is not a promise, delivery guarantee, or commitment to specific timelines.
Version History
| Version | Phase | Status | Key Features |
|---|---|---|---|
| 3.x | Core Architecture | โ Complete | Binary HDV, 3-Tier Storage, LTP/Decay |
| 4.0 | Cognitive Enhancements | โ Complete | XOR Attention, Bayesian LTP, Gap Detection, Immunology |
| 4.1 | Observability | โ Complete | Prometheus metrics, distributed tracing, project isolation |
| 4.2 | Stability | โ Complete | Async lock fixes, test suite hardening |
| 4.3 | Temporal Recall | โ Complete | Episodic chaining, chrono-weighting, sequential context |
| 5.x | The Perfect Brain | ๐ฎ Planned | Multi-Modal, Emotional, Working Memory |
Phase 5.x: The Perfect Brain
Vision: Transform MnemoCore from a sophisticated memory storage system into a truly cognitive architecture that functions as an artificial brain - but better.
5.0 Multi-Modal Memory
Goal: Enable storage and retrieval of images, audio, code structures, and cross-modal associations.
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ CURRENT: Text-only encoding โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ store("User reported bug") โ BinaryHDV โ
โ โ
โ FUTURE: Multi-modal encoding โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ store("Screenshot of error", image=bytes) โ CrossModalHDV โ
โ store("Voice note", audio=bytes) โ AudioHDV โ
โ bind(text_id, image_id, relation="illustrates") โ
โ โ
โ query("API error", modality="image") โ screenshot.png โ
โ query(image=bytes, modality="text") โ "Related conversation" โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Implementation Plan:
| Component | Description | Dependencies |
|---|---|---|
MultiModalEncoder |
Abstract encoder protocol | - |
CLIPEncoder |
Vision encoding via CLIP | transformers, torch |
WhisperEncoder |
Audio encoding via Whisper | openai-whisper |
CodeEncoder |
AST-aware code encoding | tree-sitter |
CrossModalBinding |
VSA operations across modalities | BinaryHDV |
New API Endpoints:
POST /store/multi - Store with multiple modalities
POST /query/cross-modal - Cross-modal semantic search
POST /bind - Bind modalities together
GET /memory/{id}/related - Get cross-modal related memories
5.1 Emotional/Affective Layer
Goal: Enable emotion-weighted memory storage, retrieval, and decay - mimicking how biological memory prioritizes emotionally significant events.
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ EMOTIONAL DIMENSIONS โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ โ
โ Valence: [-1.0 โโโโโโโโโโโโโโโโ +1.0] โ
โ (negative/unpleasant) (positive/pleasant) โ
โ โ
โ Arousal: [0.0 โโโโโโโโโโโโโโโโโโ 1.0] โ
โ (calm/neutral) (intense/urgent) โ
โ โ
โ EFFECT ON MEMORY: โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ High Arousal + Negative = "Flashbulb memory" (never forget) โ
โ High Arousal + Positive = Strong consolidation โ
โ Low Arousal = Faster decay (forgettable) โ
โ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
MemoryNode Extensions:
@dataclass
class MemoryNode:
# ... existing fields ...
# Phase 5.1: Emotional tagging
emotional_valence: float = 0.0 # -1.0 (negative) to +1.0 (positive)
emotional_arousal: float = 0.0 # 0.0 (calm) to 1.0 (intense)
emotional_tags: List[str] = field(default_factory=list) # ["frustration", "joy", "urgency"]
def emotional_weight(self) -> float:
"""Calculate memory importance based on emotional factors."""
# Arousal amplifies retention regardless of valence
# High arousal creates "flashbulb memories"
return abs(self.emotional_valence) * self.emotional_arousal
Modified LTP Formula:
S = I ร log(1+A) ร e^(-ฮปT) ร (1 + E)
Where E = emotional_weight() โ [0, 1]
Use Cases:
- B2B outreach: "Customer was almost in tears when we fixed their issue" โ HIGH priority
- Support tickets: "User furious about data loss" โ Never forget, prioritize retrieval
- Positive feedback: "User loved the new feature" โ Moderate retention
5.2 Working Memory Layer
Goal: Active cognitive workspace for goal-directed reasoning, not just passive storage.
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ COGNITIVE ARCHITECTURE โ
โ โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ โ WORKING MEMORY (Active) โ โ
โ โ Capacity: 7 ยฑ 2 items โ โ
โ โ โโโโโโโโโโโ โโโโโโโโโโโ โโโโโโโโโโโ โโโโโโโโโโโ โ โ
โ โ โ Goal โ โ Context โ โ Focus โ โ Hold โ โ โ
โ โ โ โ โ โ โ โ โ โ โ โ
โ โ โโโโโโโโโโโ โโโโโโโโโโโ โโโโโโโโโโโ โโโโโโโโโโโ โ โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ โ โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ โ HOT TIER (Fast Access) โ โ
โ โ ~2,000 memories, <1ms access โ โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ โ โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ โ WARM TIER (Qdrant/Redis) โ โ
โ โ ~100,000 memories, <10ms access โ โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ โ โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ โ COLD TIER (Archive) โ โ
โ โ Unlimited, <100ms access โ โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Working Memory API:
# Create working memory instance
wm = engine.working_memory(capacity=7)
# Set active goal
wm.set_goal("Troubleshoot authentication error")
# Load relevant context
wm.focus_on(await engine.query("auth error", top_k=5))
# Hold important constraints
wm.hold("User is on deadline - prioritize speed over elegance")
# Query with working memory context
results = wm.query("related issues")
# Results are RE-RANKED based on current goal + focus + held items
# Get context summary for LLM
context = wm.context_summary()
# โ "Working on: auth troubleshooting
# Focus: Recent OAuth errors
# Constraint: Time pressure"
Implementation Components:
| Component | Description |
|---|---|
WorkingMemory |
Active workspace class |
GoalContext |
Goal tracking and binding |
FocusBuffer |
Currently attended items |
HoldBuffer |
Constraints and important facts |
ContextualQuery |
Goal-directed retrieval |
5.3 Multi-Agent / Collaborative Memory
Goal: Enable memory sharing between agents while maintaining provenance and privacy.
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ COLLABORATIVE MEMORY โ
โ โ
โ Agent A Shared Memory Agent B โ
โ โโโโโโโโโโ โโโโโโโโโโโโโโโโ โโโโโโโโโโ โ
โ โ Privateโ โ โ โ Privateโ โ
โ โ Memory โโโโโโโบโ Consensus โโโโโโโโโบโ Memory โ โ
โ โโโโโโโโโโ โ Layer โ โโโโโโโโโโ โ
โ โ โ โ
โ Agent C โ Provenance โ Agent D โ
โ โโโโโโโโโโ โ Tracking โ โโโโโโโโโโ โ
โ โ Privateโโโโโโโบโ โโโโโโโโโบโ Privateโ โ
โ โ Memory โ โ Privacy โ โ Memory โ โ
โ โโโโโโโโโโ โ Filtering โ โโโโโโโโโโ โ
โ โโโโโโโโโโโโโโโโ โ
โ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Features:
- Memory provenance: Track which agent created/modified each memory
- Privacy levels: Private, shared-with-group, public
- Conflict resolution: When agents disagree on facts
- Collective intelligence: Aggregate insights across agents
5.4 Continual Learning
Goal: Enable online adaptation without catastrophic forgetting.
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ CONTINUAL LEARNING โ
โ โ
โ Traditional ML: Train โ Deploy โ (forget) โ Retrain โ
โ โ
โ MnemoCore 5.4: Learn โ Consolidate โ Adapt โ Learn โ ... โ
โ โ______________| โ
โ โ
โ KEY MECHANISMS: โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ โข Elastic Weight Consolidation (EWC) for encoder โ
โ โข Replay-based consolidation during "sleep" cycles โ
โ โข Progressive neural networks for new domains โ
โ โข Meta-learning for rapid adaptation โ
โ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Integration Priorities
Agent Frameworks
| Framework | Priority | Use Case |
|---|---|---|
| Open Claw | โญโญโญโญโญ | Primary use case, deep integration |
| LangChain | โญโญโญโญ | Memory provider plugin |
| CrewAI | โญโญโญโญ | Shared memory between agents |
| AutoGen | โญโญโญ | Conversation memory backend |
| LlamaIndex | โญโญโญ | Vector store adapter |
AI Platforms
| Platform | Priority | Integration Type |
|---|---|---|
| Claude (Anthropic) | โญโญโญโญโญ | MCP server (existing) |
| OpenAI Codex | โญโญโญโญโญ | API + function calling |
| Ollama | โญโญโญโญ | Native memory backend |
| LM Studio | โญโญโญ | Plugin architecture |
| Gemini | โญโญโญ | API adapter |
Research Opportunities
Academic Collaborations
| Area | Institutions | Relevance |
|---|---|---|
| Hyperdimensional Computing | Stanford, IBM Research, Redwood Center | Core HDC/VSA theory |
| Computational Neuroscience | MIT, UCL, KTH | Biological validation |
| Cognitive Architecture | Carnegie Mellon, University of Michigan | SOAR/ACT-R comparison |
| Neuromorphic Computing | Intel Labs, ETH Zรผrich | Hardware acceleration |
Publication Opportunities
- "Binary HDC for Long-term AI Memory" - Novel approach to persistent memory
- "Episodic Chaining in Vector Memory Systems" - Phase 4.3 temporal features
- "XOR Attention Masking for Memory Isolation" - Project isolation innovation
- "Bayesian LTP in Artificial Memory Systems" - Biological plausibility
Known Gaps (Current Beta)
- Query path is still primarily HOT-tier-centric in current engine behavior.
- Some consolidation pathways are partial or under active refinement.
- Certain integrations (LLM/Nightlab) are intentionally marked as TODO.
- Distributed-scale behavior from long-form blueprints is not fully productized.
Near-Term Priorities (Pre-5.0)
- Improve cross-tier retrieval consistency.
- Harden consolidation and archival flow.
- Improve deletion semantics and API consistency.
- Expand tests around degraded dependency modes (Redis/Qdrant outages).
- Stabilize API contracts and publish versioned compatibility notes.
- MCP server integration for agent tool access.
Not a Commitment
Items above are directional only. Order, scope, and implementation details can change during development.
Last Updated: 2025-02-18 Current Version: 4.3.0