Zep Graphiti MCP alternatives
Zep's Graphiti MCP server ships in the open-source getzep/graphiti repository, the temporal knowledge graph library under Zep's commercial service. The agent adds episodes, an LLM extracts entities and the facts between them, and every fact carries validity timestamps, so a newer episode that contradicts an older fact marks it invalid. Search combines semantic similarity, BM25 and graph distance.
The ideas are good, and Graphiti handles time better than most of the field. The cost is operational: you run FalkorDB or Neo4j, supply LLM and embedding keys, and accept a server the README calls a reference implementation with no authentication. If you want what Graphiti does well, memory that knows a decision was replaced, without operating a graph database, Glen is the first alternative to weigh.
The 9 best alternatives
- GlenOur pick
Temporal reconciliation, without running a graph database
Graphiti handles changing facts well, but you run the graph database, the LLM keys and the missing auth yourself. Glen gives you the same reconciliation as a managed service, built from every session instead of episodes the agent adds.
Recorded automaticallyClaude CodeCodexCursorSlackGitHubLinearNotionCall transcriptsOne memory for your orgSourced, permissioned per fact, newer decisions winEvery teammate's agentClaude CodeCodexCursorGets one cited answer, injected into the session before it goes looking.29%fewer tokens across 100 replayed PRs21%faster, measured end to endParity+quality on withheld testsZep GraphitiGlenWho writes memoryThe agent adds episodesRecorded from every session and toolWho can read itAnyone who reaches the portYour whole org, permissions per factWhat the agent getsFacts and nodes to assembleOne cited answer, injected in-sessionWhat you runFalkorDB or Neo4j, LLM keysNothing: hosted, OAuth 2.1 The closest to Glen in mechanism: claude-mem records sessions through hooks in Claude Code, Codex, Cursor and others, compresses tool use into observations with an observer model, and injects context when a session starts. The difference is scope. Memory lives in SQLite on one machine, grouped by project, and TeamBrain, its sharing feature, was a pilot with no public pricing when we checked. It is good memory for one developer, and your teammates' agents never see any of it.
Set up claude-mem →Mem0 runs an LLM over whatever the agent passes to add_memory, extracts discrete memories, and returns them through semantic search from a hosted endpoint at mcp.mem0.ai. It is built for per-user memory inside an application: every memory hangs off a user_id, agent_id or run_id, nothing is written unless the agent decides to call add_memory, and search hands back matching memories rather than an answer with sources. The self-run mem0-mcp package is archived and Mem0's OpenMemory was sunset in July 2026, so the hosted server is the only maintained route.
Set up Mem0 →MemPalace keeps conversations and files verbatim, without summarizing, and files each piece into wings and rooms so a search can be scoped to one person or project; ChromaDB with local embeddings does retrieval, and a SQLite graph tracks facts with validity windows. Hooks for Claude Code, Codex and Cursor save sessions automatically. A palace lives on one machine unless you run your own HTTP hub, search returns raw drawers for the agent to read rather than an answer, and its 47 tool definitions sit in every session's context window.
Set up MemPalace →Cognee runs an LLM pipeline over what it is given, extracting entities and relationships into a knowledge graph beside vector embeddings, and recall picks among graph, RAG, temporal and Cypher strategies. It runs locally on SQLite, LanceDB and an embedded graph and needs your own LLM key. Each MCP client writes to its own dataset by default, so Claude Code and Cursor on one machine don't share memory unless you change COGNEE_MCP_AGENT_SCOPED, and nothing is stored until the agent calls remember.
Set up Cognee →Supermemory's hosted server turns content the agent saves into documents, extracts memories from them asynchronously, and returns matches with a similarity score and the source document's ID. Spaces partition memory per project or user, and clients with MCP Apps support get a space picker and an interactive memory graph. Through this server, memory enters when the agent or the user calls a save tool, so what a space holds is whatever someone remembered to store.
Set up Supermemory →ByteRover keeps project knowledge as topic files in a .brv/context-tree directory inside the repository, with git-style branches and merges and optional cloud sync for a team. Its MCP server is deliberately small: brv-query runs BM25 over the tree and brv-curate saves a topic the agent wrote. Knowledge enters only when an agent or a person curates it, so the tree holds what someone chose to document, which is the AGENTS.md problem with better tooling.
Set up ByteRover →LongMemory, the CaviraOSS project published until recently as OpenMemory and unrelated to Mem0's product of that name, sorts memories into five sectors that decay at different rates and links them through a waypoint graph. It is six tools over a local SQLite file, with a user_id on each call to keep users apart. It records only what the agent chooses to store, and the team-aware rewrite on its main branch was not on npm when we checked.
Set up LongMemory (formerly OpenMemory) →Basic Memory writes memory as plain Markdown in a folder you choose, parses observations and wiki-link relations into a graph, and indexes it in SQLite for hybrid search. The files open in Obsidian and diff cleanly in git, and Basic Memory Cloud serves the same 17 tools over OAuth with shared workspaces. A note exists only when the agent calls bm_write or bm_edit, so the memory is exactly as complete as the agent's judgment about what was worth writing down.
Set up Basic Memory →
How to choose
If you want a temporal knowledge graph as a building block in your own system and are prepared to run it, Graphiti is the most rigorous open option. If you want a team's agents to get current, reconciled context with nothing to operate, Glen covers that directly. Cognee is the other graph engine to self-host; Mem0 and Supermemory are hosted memory APIs; claude-mem and MemPalace record sessions on one machine.
FAQ
- Is the Graphiti MCP server the same as Zep Cloud?
- No. The MCP server in getzep/graphiti runs the open-source library against a graph database you operate. Zep Cloud is Zep's separate hosted product, and this server does not connect to it.
- Does the Graphiti MCP server have authentication?
- Its README describes it as a reference implementation without authentication or authorization, so any client that can reach it can read or delete the graph. Put it behind your own auth before exposing it.
- Does Glen handle facts that change over time?
- Yes. When a new decision contradicts an older one, the older one stops surfacing in recall, and wrong evidence can be removed with glen_forget. Glen does not expose a graph; it returns a synthesized answer with sources.