Four Claude memory tools compared by LoreConvo's builder

The developer behind LoreConvo has published a four-way comparison of Claude memory tools, pricing his own product at $8 per month for Pro against a free tier capped at fifty sessions. The write-up, published by Labyrinthanalyticsconsulting, sets LoreConvo against Anthropic's Claude Memory Primitive, claude-mem and mem0, and names where each rival wins.

At a glance

• Storage location is the dividing line: the Memory Primitive keeps session snippets in a cloud graph tied to an API key, while claude-mem writes a client-side JSON file and syncs it to a remote store.
• Billing models diverge: Memory Primitive reads and writes ride on Claude API fees, claude-mem prices backup per gigabyte, and mem0 is free but requires a hosted or GPU-backed embedding model.
• LoreConvo serves one SQLite file over MCP to Claude Code, OpenAI Codex, Cursor and Hermes Agent, with FTS5 full-text search covering recall and no per-call fee on top.

The comparison reads as a vendor document with unusual candour, since its author credits claude-mem's 45,000 GitHub stars and concedes vector-search depth to mem0. What it exposes is the cost question underneath agent workflows: cloud-native memory bills by activity, while a local file bills by subscription. For teams whose session count grows faster than their budget, that split likely matters more than any retrieval feature list.

Memory Primitive reads and writes are billed as Claude API traffic

The Claude Memory Primitive stores session snippets in a cloud-hosted graph run by Anthropic, with each record tied to the calling API key. There is no separate memory charge, but reads and writes are API traffic, so a team running hundreds of sessions a day accumulates cost that a solo laptop user may never notice.

Data is encrypted in transit and at rest, and Anthropic retains a copy. The comparison calls that retention a deal-breaker for teams under strict data-handling policies, naming proprietary code, regulated datasets and internal architecture decisions as the cases where it introduces compliance risk.

claude-mem syncs to the cloud while mem0 hands the embedding bill to the user

claude-mem converts the graph into a JSON file written to a local directory, then periodically syncs it back to a remote store for backup. Installation is free and the optional backup service is priced per gigabyte, so a few megabytes cost almost nothing while archived weeks of history scale linearly.

Disabling the sync restores full locality but removes the automatic durability, a choice the comparison says has to be made consciously. The author credits the project's 45,000 GitHub stars and calls it the most polished community option for developers who prefer not to think about the memory layer.

mem0 builds a vector store over a local database, indexing each interaction with an embedding model and exposing add, query and delete operations. The library is free, but a hosted embedding API runs a few cents per thousand tokens and a local model consumes GPU capacity; it also enforces no access controls, leaving permissions to the operating system.

LoreConvo puts fifty free sessions and an $8 Pro tier behind one SQLite file

All LoreConvo session data lands in a single SQLite file the user owns, which can be moved, backed up or versioned with existing tooling and carries no hidden cloud component. The MCP server exposes that file to Claude Code, OpenAI Codex, Cursor and Hermes Agent, discovering a .mcp.json dropped in the project root without per-client configuration.

The free tier allows fifty sessions. Pro, at eight dollars per month, removes that limit and adds semantic search, related-session discovery and team memory sharing, all running against the same local file, so no per-call fee is layered on top of the subscription.

FTS5 full-text search handles recall without an embedding layer, while session linking, project tagging and the auto-load hook build context chains across runs. Combined with skill history tracking, the author presents the store as a lightweight audit trail of which decisions drove a pipeline design and which schemas changed.

What the auto-save does not promise

The auto-save hook runs at the end of each session and extracts a heuristic summary of decisions, tech-stack facts and open questions when the session carries enough signal, which the author describes as best-effort rather than a guarantee on every save. Pro adds JSON export and import between teammates without a central server, and the Python fallback script save_to_loreconvo.py lets automation read and write memories without registering an MCP tool.