ROADMAP.md

Roadmap

This is agentmemory-python's public roadmap. Items shift as priorities change.

• Shipped — in main
• Active — has an open PR
• Planned — accepted, not started
• Candidate — under consideration

---

Phase 1 — Foundation (done)

• Python + Flask server replacing Node.js / iii-engine
• SQLite WAL backend replacing Dolt SQL
• Audit log replacing Dolt git versioning
• BM25 search with Porter stemmer
• Gemini 768-dim vector search + hybrid RRF fusion
• 16-tool MCP endpoint
• WebSocket live stream
• Built-in HTML viewer (Dashboard, Sessions, Memories, Graph, Timeline, Lessons, Slots, Replay)
• Knowledge graph visualization — folder nodes with unique colors, force-directed layout
• 4-tier memory consolidation (Working → Episodic → Semantic → Procedural)
• Lessons system with confidence decay
• HuggingFace Space deployment with sync fingerprinting
• Dolt → SQLite one-time migration (7502 rows, 19.6MB)

---

Phase 2 — Reliability

Active

• Pytest test coverage for src/functions.py core operations (observe, remember, search, context)
• Graph edge label bug — labels overlap at default zoom on dense graphs

Planned

• Additional embedding providers — OpenAI text-embedding-3-small, local sentence-transformers
• Hook scripts — prebuilt bash/PowerShell hook scripts for Claude Code, Cursor, Codex CLI pointing at http://localhost:3111
• Memory export/import — JSON round-trip so users can migrate between instances
• Health endpoint (/agentmemory/health) — richer than /livez, includes index sizes, sync status, last backup time
• Graceful shutdown — flush BM25/vector index to DB before exit on SIGTERM

---

Phase 3 — Breadth

Candidate

• Additional LLM providers — OpenRouter, Ollama (local), Cohere for compression and consolidation
• MCP tool parity — expand from 16 to 30+ tools to match the most-used subset of the Node.js 53-tool surface
• GitHub Actions hook — observe CI runs as memory events
• Slack / Discord connector — ingest messages as observations
• Multi-agent shared memory — namespace isolation for team use; agents share a pool with per-agent write attribution
• RBAC — role-based access control for shared deployments
• Benchmark harness — reproduce the LongMemEval-S R@5 metric on the Python stack
• pip package — pip install agentmemory with a agentmemory CLI entrypoint

---

Non-goals

• Full parity with the 128-endpoint / 53-tool Node.js surface — the Python version targets simplicity and HF deployment, not feature count
• iii-engine worker model — the Python version intentionally avoids the iii runtime
• Dolt git-versioned history — the audit log satisfies the write-tracking requirement without the complexity