---
type: analysis
slug: guide-agent-memory-mem0-zep-letta
title: "Agent memory: mem0, Zep, Letta"
description: "mem0, Zep, and Letta are third-party agent memory layers, unlike Claude/ChatGPT built-in memory which is product-scoped personalization inside those chat apps. mem0 extracts facts into SQL + vector embeddings (+ Platform Graph Memory entity linking), with Platform cloud or OSS self-host (pip/npm Mem"
published_at: 2026-11-13T09:00:00.000Z
topics: ["tools","developer_tools"]
source_urls: ["https://docs.mem0.ai/core-concepts/how-it-works","https://docs.mem0.ai/platform/features/graph-memory","https://docs.mem0.ai/platform/platform-vs-oss","https://docs.mem0.ai/platform/mem0-mcp","https://docs.mem0.ai/open-source/overview","https://help.getzep.com/v4/concepts.md","https://help.getzep.com/v4/facts.md","https://help.getzep.com/v4/context-mcp-server.md","https://help.getzep.com/v4/zep-vs-graphiti.md","https://docs.letta.com/concepts/memfs/index.md","https://docs.letta.com/agent-sdk/memory/index.md","https://docs.letta.com/platform/hosted-mcp/index.md","https://docs.letta.com/self-hosting/index.md","https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context","https://help.openai.com/en/articles/8590148-memory-in-chatgpt","https://mem0.ai/pricing","https://www.getzep.com/pricing","https://docs.mem0.ai/platform/pricing","https://docs.mem0.ai/changelog/highlights"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
mem0, Zep, and Letta are third-party agent memory layers, unlike Claude/ChatGPT built-in memory which is product-scoped personalization inside those chat apps. mem0 extracts facts into SQL + vector embeddings (+ Platform Graph Memory entity linking), with Platform cloud or OSS self-host (pip/npm Memory). Zep stores temporal Context Graphs (bi-temporal facts) in a managed Context Lake (or OSS Graphiti), with Python/TS/Go SDKs and Context MCP. Letta agents keep git-backed MemFS memory (labeled Markdown files in context), Cloud or fully local App Server/CLI; MCP at api.letta.com/mcp creates/messages agents. Claude saves individual topics; ChatGPT uses saved memories plus optional chat-history reference—neither exposes a developer memory API comparable to these SDKs.

*Field guide compiled 2026-11-06 from the official documentation, pricing pages and changelogs linked below; every line links to the page that states it.*

## What it does

- **How Mem0 Works.** Mem0 sits between app and model: add() extracts durable facts (not full transcripts by default), embeds them, and stores facts/metadata in SQL, embeddings in a vector DB, and extracted entities in an entity store; search() ranks with semantic, keyword, entity, and temporal signals. Scope with userid/agentid/runid filters. How: Use MemoryClient (Platform) or Memory (OSS); call add after useful turns and search before the next model call. [source](https://docs.mem0.ai/core-concepts/how-it-works)
- **Graph Memory.** Mem0 Platform builds a native graph: entities from memories become nodes; memories sharing an entity are linked; search boosts entity-connected memories into the combined score. Always on for retrieval on all plans; interactive Graph view is Pro/Enterprise. No external Neo4j/Memgraph to provision. OSS no longer has queryable graph memory (external graph stores removed). [source](https://docs.mem0.ai/platform/features/graph-memory)
- **Platform vs Open Source.** Same core add/search/update/delete loop. Platform is managed (hosted vector/LLM/embedder, dashboard, Graph Memory, Memory Decay, Temporal Reasoning, Dream). OSS self-hosts: you configure vector DB, LLM, embedder; entity boost remains but no queryable graph, decay, temporal params, or Dream. [source](https://docs.mem0.ai/platform/platform-vs-oss)
- **Mem0 MCP.** Hosted MCP server at https://mcp.mem0.ai/mcp exposes addmemory, searchmemories, get/update/delete tools to Claude, Claude Code, Cursor, Codex, Windsurf, VS Code, OpenCode. Requires Platform account/API key; memories live in the Mem0 account. How: npx mcp-add --url https://mcp.mem0.ai/mcp or manual HTTP MCP config; browser OAuth or bearer API key. [source](https://docs.mem0.ai/platform/mem0-mcp)
- **Open Source Overview.** Self-host as library (Memory()) or Docker REST server. Library defaults: OpenAI gpt-5-mini LLM, text-embedding-3-small, local Qdrant at /tmp/qdrant, SQLite history. Server defaults: Postgres+pgvector. Many vector stores/LLMs/embedders configurable. [source](https://docs.mem0.ai/open-source/overview)
- **Context Graph / Context Lake.** Zep is a unified context layer: temporal Context Graphs hold facts, entities, episodes, thread summaries, observations, and user summary; the Context Lake manages many graphs with governance. Agent memory uses a per-user graph + threads; retrieve via thread.getcontext (Context Block) or graph search tools. [source](https://help.getzep.com/v4/concepts.md)
- **Facts.** Zep facts are time-stamped relationships on edges with bi-temporal timestamps: validat/invalidat (world timeline) plus createdat/expiredat (graph timeline). New data can invalidate prior facts; Context Block includes date ranges for point-in-time answers. [source](https://help.getzep.com/v4/facts.md)
- **Context MCP Server.** Endpoint https://api.getzep.com/mcp lets end users connect Claude/ChatGPT/Cursor/Codex etc. to their own Zep user graph via IdP sign-in. Tools include searchgraph, getusersummary, getsubgraph, addmemory, and shared-graph variants. Uses per-account MCP seats; Custom OIDC needs Enterprise. [source](https://help.getzep.com/v4/context-mcp-server.md)
- **Zep vs Graphiti.** Graphiti is Zep's open-source temporal knowledge-graph framework (pluggable Neo4j/FalkorDB/Neptune, hybrid vector+full-text+graph retrieval, self-hosted MCP). Zep Cloud uses Graphiti inside a managed Context Lake on Konig with Observations, SDKs, governance, SOC 2/HIPAA, Cloud/BYOK/BYOC. [source](https://help.getzep.com/v4/zep-vs-graphiti.md)
- **MemFS.** Letta agent long-term memory is a git repository projected as MemFS (context repository). Root Markdown files (e.g. persona.md) load into the system prompt every turn; directories with MEMORY.md stay out of context until read. Commits/pushes sync memory across computers. No semantic/vector index by default; optional MemFS Search mod + QMD for hybrid search. [source](https://docs.letta.com/concepts/memfs/index.md)
- **Memory (Agent SDK).** createAgent accepts persona/human shorthand or memory{label,value} entries that become Markdown files under system/. Dreaming configures background subagents (trigger step-count/compaction; behavior reminder/auto-launch). Shared memory repositories are Cloud-only for multi-agent shared git files. [source](https://docs.letta.com/agent-sdk/memory/index.md)
- **Hosted Letta MCP server.** Any MCP client can call https://api.letta.com/mcp with a Letta API key bearer token to list/create agents and sendagentmessage / getreply. Agents keep memory across conversations. Client must support custom bearer headers (OAuth-only clients not supported yet). [source](https://docs.letta.com/platform/hosted-mcp/index.md)
- **Self-hosting.** Local runtime: CLI --backend local, state on-device, no Letta account required. Self-hosted App Server: letta server --backend local --listen; Agent SDK backends local or remote. Self-hosted agents are not on chat.letta.com and need manual backup. State under ~/.letta/lc-local-backend including memfs/agent-id/memory. [source](https://docs.letta.com/self-hosting/index.md)
- **Claude memory (topics).** Claude saves memory as individual topics while you chat (not only post-hoc daily summaries on the new experience). On by default for Free/Pro/Max/Ultra; Team/Enterprise owner-gated. Project memory is separate; chat search (RAG) is a distinct paid-plan feature. Not a third-party developer memory SDK. [source](https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context)
- **Memory in ChatGPT.** When enabled, ChatGPT personalizes from past chats, saved memories, custom instructions, Library/Space files, and connected apps. Saved memories are separate from chat history; Reference chat history can be toggled (off schedules deletion of chat-derived memory within 30 days). Controls under Settings Personalization Memory. Product personalization, not an external agent memory API. [source](https://help.openai.com/en/articles/8590148-memory-in-chatgpt)

## Modes

- **Mem0 Platform (managed).** Hosted API MemoryClient; Graph Memory, Decay, Temporal Reasoning, Dream on Platform; MCP at mcp.mem0.ai. [source](https://docs.mem0.ai/platform/platform-vs-oss)
- **Mem0 Open Source (self-hosted).** Library Memory() or Docker server; bring your own vector store/LLM/embedder; no Platform-only ranking features. [source](https://docs.mem0.ai/open-source/overview)
- **Zep Cloud / Cloud+BYOK / BYOC.** Managed Cloud; Cloud with customer KMS keys; or Bring Your Own Cloud VPC. Community Edition self-host deprecated. [source](https://www.getzep.com/pricing)
- **Graphiti (open source).** Self-host temporal Context Graphs with Neo4j/FalkorDB/Neptune; experimental Graphiti MCP server. [source](https://help.getzep.com/v4/zep-vs-graphiti.md)
- **Letta local / self-hosted App Server.** CLI --backend local or letta server --backend local; state under ~/.letta; no automatic cloud backup; not accessible via chat.letta.com. [source](https://docs.letta.com/self-hosting/index.md)

## Pricing and limits

- **Mem0 Free:** USD $0. Unlimited end users. First 30 days: 10,000 adds and 1,000 retrievals/month with model calls on Mem0. With your own LLM key: 50,000 adds and 10,000 retrievals/month. Usage-based billing past limits with a monthly cap you set. Community support. [source](https://mem0.ai/pricing)
- **Zep Free:** USD $0. 10,000 credits/month; no rollover or auto-topup; 2 projects; 1 Context MCP Server seat; 5 custom entity & edge types; variable rate limits; lower priority Episode processing; feature/service levels may change. [source](https://www.getzep.com/pricing)
- **Zep Flex:** USD $125/month billed monthly (annual option saves 17%). 50,000 credits/month then $25/10,000 credits; auto top-up at 20%; 30-day rollover; 600 RPM; 5 projects; 5 Context MCP seats; 10 custom entity/edge types; API logs 1 day; unlimited memories/retrieval/users. Credits: 1 per Episode up to 350 bytes (+1 per additional 350 bytes); 1/8 credit per webhook; 0 for retrieval/storage/users/graph storage. [source](https://www.getzep.com/pricing)
- **Zep Flex Plus:** USD $375/month billed monthly. 200,000 credits/month then $75/40,000 credits; 60-day rollover; 1,000 RPM; 10 projects; 15 Context MCP seats; Observations; custom extraction; webhooks; analytics; API logs 7 days. [source](https://www.getzep.com/pricing)

## Recent changes

- 2026-11-05: Zep announced Agent Skills generally available (blog dated Nov 05, 2026); Skills free on all plans including Free until January 4, 2027. [source](https://www.getzep.com/pricing)
- 2026-10-26: Mem0 pricing change: new accounts Free + Custom; pre-Oct 26 Starter/Growth/Pro grandfathered. Free 30-day window for pre-change signups ends November 25, 2026. [source](https://docs.mem0.ai/platform/pricing)
- 2026-05-13: Mem0 Platform v3 Temporal Reasoning for time-aware search ranking (last week/upcoming/as of…). [source](https://docs.mem0.ai/changelog/highlights)
- 2026-05-08: Mem0 Memory Decay: opt-in per-project search-time boost for recently used memories (0.3×–1.5× soft bias). [source](https://docs.mem0.ai/changelog/highlights)
- 2026-04-14: Mem0 SDK/platform new memory algorithm (v2/v3): ADD-only extraction, hybrid retrieval, built-in Graph Memory; external graph stores removed from OSS. [source](https://docs.mem0.ai/changelog/highlights)

## Alternatives

- **Graphiti vs Zep.** Choose Graphiti for flexible OSS graph core you operate; choose Zep for managed enterprise Context Lake, governance, Observations, and support. [source](https://help.getzep.com/v4/zep-vs-graphiti.md)
- **Mem0 Platform vs OSS.** Platform for zero infra + Graph Memory/Dream/Decay/Temporal; OSS for full data control and custom providers without Platform-only features. [source](https://docs.mem0.ai/platform/platform-vs-oss)

## Traps

- Mem0 Free after 30 days without a BYOK LLM key: memories remain and searches work up to 1,000/month, but new adds error until a key is connected. Past monthly limits, calls error until reset. [source](https://docs.mem0.ai/platform/pricing)
- Letta self-hosted agents cannot be opened in chat.letta.com and are not auto-backed-up; CLI cannot connect to a remote App Server URL (SSH to that machine instead). Agent SDK is TypeScript-only (no Python Agent SDK). [source](https://docs.letta.com/self-hosting/index.md)
- Claude: deleting a chat does not remove memory entries generated from it; reset memory is permanent. Sensitive topics off by default; some IDs/financial/criminal/immigration data never saved even if asked. [source](https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context)
- ChatGPT: deleting a chat does not necessarily delete a separate saved memory from that chat. Turning Reference chat history off schedules deletion of chat-derived remembered info within 30 days (original chats remain). Memory summary is high-level and may omit details. [source](https://help.openai.com/en/articles/8590148-memory-in-chatgpt)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Tools](https://newruntime.com/topics/tools.md): Explore the tools topic hub.
- topic hub: [Developer Tools](https://newruntime.com/topics/developer-tools.md): Explore the developer_tools topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
