Skip to content

ROADMAP.md

Latest commit

 

History

History
186 lines (145 loc) · 7.07 KB

ROADMAP.md

File metadata and controls

186 lines (145 loc) · 7.07 KB

Omnimind Development Roadmap

Phase 1: Core Memory Engine

Status: COMPLETE ✅

Goal: A fast, local memory store with semantic search.

Delivered

  • Project scaffolding (TypeScript strict, Vitest, build system)
  • SQLite unified schema (vectors, text, graph, metadata, activity log)
  • SQLite + sqlite-vss integration (ARM64 native build)
  • EmbeddingEngine (ONNX all-MiniLM-L6-v2, 384d) — runs locally, no Python
  • MemoryStore CRUD with Result<T,E> pattern
  • Semantic search (vector similarity via sqlite-vss)
  • Keyword search (FTS5)
  • Hybrid search (vector + keyword + graph fusion with reciprocal rank)
  • Temporal queries (time-range filtering + temporal half-life boost)
  • Turn-level storage (storeTurns()) with batch embedding and shared sourceId
  • L0: VerbatimLayer — full text, immediate storage
  • L1: CompressedLayer — AAAK-style shorthand compression
  • L2: ConceptGraph — entity-relationship extraction + graph persistence
  • L3: WisdomLayer — pattern distillation
  • AgingPipeline — lazy transitions between layers (triggered on access)
  • Pin system — prevent aging for important memories
  • Import/Export — lossless JSON + human-readable Markdown

Benchmarks

  • Retrieval R@5: 94% on LongMemEval-S (turn-level) / 84% (session-level)
  • Retrieval latency: < 20ms (p95)
  • Test coverage: 135 tests passing

Phase 2: Proactive Prediction

Status: COMPLETE ✅

Goal: Memory that predicts what you need before you ask.

Delivered

  • ActivityTracker — file change monitoring, tool invocation history, time patterns
  • Context fingerprint construction (project, branch, file, time, recent tools)
  • IntentPredictor — exact + similar signature matching, frequency/recency scoring
  • Memory relevance scoring with confidence threshold (70% cutoff)
  • ContextInjector — pre-load predictions into MCP context
  • PatternStore with SQLite persistence

Benchmarks

  • Top-1 accuracy: 73.9% (target: ≥70%) ✅
  • Top-3 accuracy: 91.3%
  • Wake-up tokens: < 150 (context injection is lightweight)

Phase 3: Cross-Tool Memory Bus

Status: IN PROGRESS 🔄

Goal: One memory for Claude, ChatGPT, Cursor, and any MCP-compatible tool.

Delivered

  • MemoryBus core — pub/sub event routing with vector clocks
  • ConflictResolver — tool-priority tiebreaking + supersede resolution
  • Memory event schema (create, update, delete, access, sync_request)
  • Dead-letter queue for failed routing
  • Claude Code adapter — file watch on ~/.claude/projects, auto-save turns from .jsonl, decision extraction, CLAUDE.md notifications
  • Cursor adapter — file watch on ~/.cursor/conversations, supports .jsonl + .json formats, auto-save turns
  • ChatGPT adapter — file watch on ~/.chatgpt/exports, parses official OpenAI export format (message tree → flat turns), multi-part content support
  • MemoryBus.storeTurns() — direct batch storage bypassing event pipeline

Remaining

  • Cross-device sync protocol (encrypted P2P)
  • Sync conflict UI indicators
  • Memory merge strategies UI (automatic + manual)

Deliverable

Memories created in Claude Code appear in Cursor automatically via shared MemoryStore.

Phase 4: Visual Memory Explorer

Status: IN PROGRESS 🔄

Goal: A GUI that lets you see, explore, and manage your memory.

Delivered

  • Tauri desktop shell + Svelte 5 frontend
  • IPC bridge (Node.js sidecar server on localhost:8844)
  • Search interface — natural language queries with hybrid results
  • Temporal timeline — chronological memory view
  • Concept graph — D3.js force-directed entity-relationship visualization
  • Spatial memory map — 2D wing/room layout (memory palace), zoom/pan, layer-colored dots, detail sidebar
  • Import/export UI (JSON + Markdown)
  • Settings panel — theme, auto-start, configuration

Remaining

  • Drag-and-drop card repositioning in spatial map
  • Connection lines between related memories in spatial map
  • Manual memory editing (double-click to edit)
  • Real-time predictions panel in GUI

Deliverable

Cross-platform desktop app (macOS primary, Linux/Windows via Tauri).

Phase 5: MCP Server & Integration

Status: IN PROGRESS 🔄

Goal: Seamless integration with any MCP-compatible AI tool.

Delivered

  • Full MCP server implementation (@modelcontextprotocol/sdk)
  • Tool definitions with Zod schemas:
    • omnimind_search — semantic search across all layers
    • omnimind_store — store new memory with auto-layering
    • omnimind_store_conversation — batch store conversation turns
    • omnimind_predict — get predicted memories for current context
    • omnimind_status — system health and layer statistics
    • omnimind_subscribe — subscribe to memory events
    • omnimind_sync — sync with other tools
  • Auto-save hooks for Claude Code (turn-level file watching)
  • Context injection endpoint (/api/context)

Remaining

  • Multi-agent memory isolation (each agent gets its own namespace/wing prefix)
  • Memory-aware context compression (truncate while preserving Omnimind context)
  • Prompt templates (memory-aware system prompts)
  • Resource exposure (memory as readable MCP resources)
  • Published npm package with npx omnimind one-liner setup

Future Phases (Post-MVP)

Phase 6: Team Memory (Month 4)

  • Multi-user memory spaces with access controls
  • Shared project memories
  • Memory permissions (read/write/admin)
  • Audit logging

Phase 7: Advanced Neuro-Symbolic (Month 5)

  • Latent concept space (beyond embeddings)
  • Cross-modal memory (code + docs + conversations)
  • Memory inference (deduce new facts from stored ones)
  • Analogical reasoning ("this situation is like that past one")

Phase 8: Memory Marketplace (Month 6)

  • Import domain knowledge packs
  • Share memory templates
  • Community-curated wisdom libraries

Success Metrics

Metric Current Phase 5 (MVP) Phase 8
Retrieval R@5 94% >= 97% >= 98%
Retrieval latency < 20ms < 15ms < 10ms
Prediction accuracy 82.6% top-1 >= 75% >= 85%
Wake-up tokens < 150 < 150 < 100
Storage growth Linear Sub-linear Sub-linear
Tools supported 3 (Claude, Cursor, ChatGPT) 3+ 5+
Test coverage 135 tests >= 85% >= 90%

Current Sprint

Sprint: MCP Polish + Multi-Agent Isolation

Priority: Multi-agent namespace isolation + context compression

Blocked by: None Blocking: npm package publish

Focus areas:

  1. Multi-agent memory isolation (namespace prefixes per agent)
  2. Context compression that preserves Omnimind predictions
  3. npm package preparation (npx omnimind setup)
  4. Documentation update (README, API docs, integration guides)

Definition of done:

  • Each MCP client can operate in its own memory namespace
  • Context injection respects token budget and compresses gracefully
  • npm test passes with >= 85% coverage
  • Package ready for npm publish (dry-run passes)