agentcairn vs other agent-memory tools
Most agent-memory tools keep your memory in a database. agentcairn keeps it as a Markdown vault you own, where the files are the source of truth — not an export.
At a glance
| Dimension | agentcairn | basic-memory | Mem0 | Letta | Zep |
|---|---|---|---|---|---|
| Source of truth | Markdown vault | Markdown files | database | database (Postgres) | knowledge graph |
| Plain files you own/edit | Yes | Yes | No | No | No |
| Runs as | daemonless CLI + on-demand MCP | local CLI + MCP (optional cloud) | library / self-host / cloud | self-host (de-emphasized) + cloud | hosted SaaS (self-host CE deprecated) |
| Retrieval | hybrid BM25+vector (RRF) + reranker + wikilink graph | hybrid full-text+vector + wikilink graph | hybrid (semantic+BM25+entity) | pgvector-based | temporal knowledge graph |
| License | Apache-2.0 | AGPL-3.0 | Apache-2.0 | Apache-2.0 | Graphiti Apache-2.0; Zep Cloud proprietary |
| Focus | memory for coding agents | local-first knowledge/memory | memory layer for LLM apps | stateful-agents framework | enterprise agent memory |
The tools, fairly
Mem0
An Apache-2.0 open-source memory layer for LLM apps, available as an embeddable library, a self-hosted server, or a managed cloud platform. Memory lives in a vector database (not human-readable files). Hybrid retrieval (semantic + BM25 + entity). It is not cloud-only.mem0.ai·GitHub
Letta
Formerly MemGPT. An Apache-2.0 framework for stateful agents — memory is one part of a broader agent platform, not a standalone memory store. State lives in Postgres (pgvector). It ships a hosted cloud plus a self-hostable server that is officially de-emphasized.letta.com·GitHub
Zep
A context-engineering platform built on a temporal knowledge graph. The shipping product, Zep Cloud, is a proprietary hosted service. Its open-source engine, Graphiti (Apache-2.0), can be used on its own, but the self-hostable Zep Community Edition is deprecated. So “open source” applies to Graphiti, not the hosted product.getzep.com
basic-memory
agentcairn’s closest peer — local-first, with Markdown files as the source of truth and a derivative local index, plus a wikilink knowledge graph and hybrid retrieval. The differences are narrow and factual: it is AGPL-3.0 (vs agentcairn’s permissive Apache-2.0), and it is a general human+AI knowledge system, whereas agentcairn is purpose-built for coding-agent memory — multi-harness plugins (Claude Code, Codex, Cursor), out-of-band session-transcript capture, secret redaction before write, bi-temporal validity, and provenance-aware recall.GitHub·docs
When not to choose agentcairn
agentcairn is local-first memory for coding agents. It is the wrong tool when:
- You want a hosted, managed multi-tenant SaaS → Mem0 or Zep.
- You need a full stateful-agent framework, not just memory → Letta.
- You need enterprise graph memory at scale → Zep.
- You want a drop-in memory API for a general (non-coding-agent) LLM app → Mem0.
Learn more
FAQ
Is agentcairn a Mem0 alternative?
Yes — both give AI agents long-term memory, but agentcairn stores it as a local Markdown vault you own rather than in a vector database, and is built specifically for coding agents.
Can I move my memory off agentcairn?
Yes — your memory is plain Markdown files in a folder you control. There's nothing to export; the DuckDB index is a rebuildable cache, not the source of truth.
How is agentcairn different from basic-memory?
They share a local-first, Markdown-as-source-of-truth design. agentcairn is Apache-2.0 (basic-memory is AGPL-3.0) and is purpose-built for coding-agent memory — multi-harness plugins, out-of-band transcript capture, secret redaction, and provenance-aware recall.
Get started
Install agentcairn into your agent in one command:
cairn install --all