Roadmap¶
What MemoryMesh Is¶
MemoryMesh is the SQLite of AI Memory -- an embeddable, zero-dependency Python library that gives any LLM application persistent, intelligent memory.
Primary audience: AI tool users (Claude Code, Cursor, Gemini CLI, Codex CLI). Install once, your AI remembers everything.
Secondary audience: Developers building LLM apps. Three lines of Python. No servers, no infrastructure, no vendor lock-in.
v3.2 -- Launch Ready (Q1 2026)¶
Security hardening, test coverage, and distribution. Feature freeze until real user feedback is collected.
- Security fixes (metadata filter validation, file permissions, encryption documentation)
- MCP server test coverage
- README and messaging rewrite (pain-first narrative, demo GIF, separate MCP quick start)
- Documentation polish (slimmer getting-started, MCP verification step)
- Show HN launch, Reddit, awesome list submissions, MCP directory listing
v4.0 -- Invisible Memory (Shipped v4.0.1)¶
The AI shouldn't need to "use" MemoryMesh. It should just work. Status: shipped on PyPI and GitHub.
Smart Sync¶
Export the top-N most relevant memories to .md files, ranked by importance and recency -- not a full dump. Directly reduces token cost by injecting only what matters into every session.
- User-configurable ranking weights
- Graceful degradation to current full-dump behavior
Auto-Remember Hooks¶
PostToolUse and Stop hooks that capture decisions, patterns, and key facts without requiring the AI to call remember(). Zero-instruction persistence -- memory happens as a side effect of working.
- Noise filtering gate before storing (length, keyword density, dedup)
- Privacy guard runs before capture (hooks may see API keys in tool output)
Lean MCP¶
Consolidate MCP tools where real usage data supports it. Less schema overhead per session, same capabilities.
Task-Aware Injection¶
session_start reads the user's first message and generates targeted context instead of a generic profile dump.
Scaling & Architecture¶
- Fix full-table embedding scan (SQL-level pre-filter before loading vectors)
- SQLite FTS5 for keyword search when embeddings are disabled
- Split MCP server monolith into protocol + tools modules
- Custom exception hierarchy
Measured Overhead¶
Instrument real token impact per session. Prove the value with data, not claims.
v5.0 -- Adaptive Memory (2027)¶
Lightweight heuristics first. LLM-based anticipation deferred.
- Question frequency tracking -- Surface answers proactively when a topic recurs
- Behavioral patterns -- Learn coding styles and preferred approaches from access data
- Multi-device sync -- Document Syncthing/rsync with
.memorymesh/(zero code) - Cross-session linking -- Connect related sessions into chains
Completed Milestones¶
| Version | Milestone |
|---|---|
| v1.0 | Production-ready core: episodic memory, auto-importance, encrypted storage, compaction |
| v2.0 | Personality engine: 9 memory categories, auto-categorization, session_start, structured sync |
| v3.0 | Intelligent memory: pin support, privacy guard, contradiction detection, retrieval filters, web dashboard |
| v3.1 | Setup & diagnostics: improved onboarding, health checks, runtime reconfiguration |
| v4.0 | Invisible Memory: Smart Sync, configurable weights, EncryptedStore completeness, security hardening |