MemoryMesh

Option 2: Install locally (recommended for production)

Install once, then add the config to your tool of choice:

Claude Code(~/.claude/settings.json):

{ "mcpServers": { "memorymesh": { "command": "memorymesh-mcp" } } }

{ "mcpServers": { "memorymesh": { "command": "memorymesh-mcp" } } }

{ "mcpServers": { "memorymesh": { "command": "memorymesh-mcp" } } }

Your AI now has persistent memory across sessions. Preferences, decisions, and patterns survive context window resets.

from memorymesh import MemoryMesh memory = MemoryMesh() memory.remember("User prefers Python and dark mode") results = memory.recall("What does the user prefer?")

That is it. Three lines to give your AI application persistent, semantic memory.

# Works with any LLM -- inject recalled context into your prompts context = memory.recall("What do I know about this user?") # Claude response = claude_client.messages.create( model="claude-sonnet-4-20250514", system=f"User context: {context}", messages=](https://smithery.ai/servers/sparkvibe-io/memorymesh)[{"role": "user", "content": "Help me design an API"}], ) # GPT response = openai_client.chat.completions.create( model="gpt-4", messages=[ {"role": "system", "content": f"User context: {context}"}, {"role": "user", "content": "Help me design an API"}, ], ) # Or Ollama, Gemini, Mistral, Llama, or literally anything else

- Store-- After each interaction,remember()the key facts, decisions, and patterns (not the full conversation).
- Recall-- At the start of the next session,recall()retrieves only the most relevant memories ranked by semantic similarity, recency, and importance.
- Persist-- Memories live in SQLite on your machine. They survive session restarts, tool switches, and context window resets.

- Cross-session persistence-- Decisions made Monday are still known Friday.
- Cross-tool memory-- What you teach Claude stays available in Gemini, Codex, and Cursor.
- Structured recall-- Categories, importance scoring, time decay, and semantic search instead of brute-force history replay.
- Privacy-- Everything local. No cloud, no telemetry, no data leaves your machine.

# Base installation (no external dependencies, uses built-in keyword matching) pip install memorymesh # With local embeddings (sentence-transformers, runs entirely on your machine) pip install "memorymesh[local]" # With Ollama embeddings (connect to a local Ollama instance) pip install "memorymesh[ollama]" # With OpenAI embeddings pip install "memorymesh[openai]" # Everything pip install "memorymesh[all]"

- Simple API--remember(),recall(),forget(). That is the core interface. No boilerplate, no configuration ceremony.
- SQLite-Based-- All memory stored in SQLite files. No database servers, no infrastructure. Automatic schema migrations.
- Framework-Agnostic-- Works with any LLM, any framework, any architecture. Use it with LangChain, LlamaIndex, raw API calls, or your own setup.
- Pluggable Embeddings-- Choose from local models, Ollama, OpenAI, or plain keyword matching with zero dependencies.
- MCP Support-- Built-in MCP server for seamless integration with Claude Code, Cursor, Gemini CLI, and other MCP-compatible tools.
- Memory Categories-- Automatic categorization with scope routing. Preferences go global; decisions stay in the project. MemoryMesh decides where memories belong.
- Encrypted Storage-- Optionally encrypt memory text and metadata at rest with zero external dependencies.
- Privacy-First-- All data stays on your machine. No telemetry, no cloud calls, no data collection. You own your data.
- Auto-Compaction-- Transparent deduplication that runs automatically during normal use. Like SQLite's auto-vacuum, you never need to think about it.
- Cross-Platform-- Runs on Linux, macOS, and Windows. Anywhere Python runs, MemoryMesh runs.

- Bulk Access Updates--recall()batches N access-time updates into 1-2 SQL calls instead of N.
- Light Listing--session_startandsmart_syncskip loading embedding blobs, reducing I/O.
- Recency Fix--update_access()no longer setsupdated_at, fixing a recency feedback loop.

- Contradiction Scan-- 10K→500 candidate limit for contradiction detection (biggest perf win).
- Security-- CORS same-origin, 1MB body limit, SSRF blocklist expansion, MCP assert→if/raise.
- Correctness-- Atomic scope migration (save-first-then-delete),on_conflictvalidation.
- Infra--PRAGMA busy_timeout=5000, PEP 561py.typed, expanded secret regex patterns.

- Smart Sync-- Export the top-N most relevant memories to.mdfiles, ranked by importance and recency.
- Configurable Relevance Weights-- Tune recency, importance, and similarity weights via environment variables or constructor parameters.
- EncryptedStore Completeness--EncryptedMemoryStorenow supportssearch_filteredandupdate_fields, matching the fullMemoryStoreinterface.
- Security Hardening-- SQL injection fix insearch_filtered(strict allowlist for metadata keys) and explicit file permissions on database files.

v4.3.0is the latest release. Available onPyPI.

See thefull roadmapfor details, strategic context, and completed milestones.

Full documentation:sparkvibe-io.github.io/memorymesh

We welcome contributions from everyone. SeeCONTRIBUTING.mdfor guidelines on how to get started.

MIT License. SeeLICENSEfor the full text.

MemoryMesh is part of theSparkVibeopen-source AI initiative. We believe that foundational AI tools should be free, open, and accessible to everyone -- not locked behind paywalls, cloud subscriptions, or proprietary platforms.

MCP server for Apache AGE graph databases on PostgreSQL. 21 MCP tools — the most comprehensive Apache AGE MCP server (graph CRUD, Cypher queries, batch transactions, semantic search, Graph RAG, vis.js visualization, export/import) - F#/.NET — the only non-Python Apache AGE MCP server, installs as a single dotnet tool - Production-grade — BenchmarkDotNet-verified performance (cached queries in 62 ns, Cypher in 1 ms) - Open source — MIT license, published on NuGet - Documentation — full docs site at neftedollar.com/age-mcp