Skip to content

Developer tools

Four Claude memory tools compared by LoreConvo’s builder

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: labyrinthanalyticsconsulting.com

The developer behind LoreConvo, a local memory layer for Claude, published a four-way comparison of Claude memory tools on August 18. It puts his own product next to Anthropic’s Claude Memory Primitive, claude-mem and mem0. In the Labyrinth Analytics Consulting post, the four tools differ on two points: where the memory physically lives, and whether the bill grows with how often you read from it.

He discloses the conflict of interest at the start: “I have tried all four; I built LoreConvo, so I will tell you exactly where the others win.”

He describes the Claude Memory Primitive as storing session snippets in a cloud-hosted graph that Anthropic manages. You send a request, Claude returns a short summary, and the service keeps a record tied to your API key. Data is encrypted in transit and at rest, but Anthropic keeps a copy. The author argues that for teams under strict data-handling policies this retention is a deal-breaker, and that it creates compliance risk for proprietary code or regulated datasets. Memory has no separate price, but every read from or write to the graph is a billed API call. He says that is cheap for a solo developer on a laptop, while a team running hundreds of sessions a day sees the extra calls add up quickly.

Anthropic describes its own memory tool differently. In Anthropic’s write-up, which introduced the tool alongside context editing for Claude Sonnet 4.5, Claude creates, reads, updates and deletes files in a dedicated memory directory that lives in the developer’s infrastructure. Anthropic says the tool “operates entirely client-side through tool calls” and that developers manage the storage backend.

claude-mem, in the author’s account, sits on the client side and converts the graph into a local JSON file, which it periodically syncs to a remote store for backup. Installing it is free. The optional backup is priced per gigabyte stored: a few megabytes of sessions cost close to nothing, and the cost grows linearly as weeks of history pile up. You can turn the sync off, but you lose the automatic durability, and you have to remember to do it. He credits the project with its 45,000 GitHub stars and calls it “the most polished community option” for solo developers who want capture to happen while they keep working.

The claude-mem documentation describes a different storage layer: a SQLite database of sessions, observations and summaries with FTS5 search, plus an optional Chroma vector index for semantic search. Five lifecycle hooks (SessionStart, UserPromptSubmit, PostToolUse, Summary on Stop, and SessionEnd) feed a worker service that compresses tool-use observations into summaries for later sessions. For privacy, the docs list <private> tags for keeping blocks out by hand and regex-based auto-redaction for secrets. The tool needs Node.js 20.0.0 or higher. Besides Claude Code, it supports Cursor, Windsurf, OpenCode, Codex CLI, Antigravity CLI and OpenClaw.

mem0 puts the most work into retrieval. The author describes it as a vector store on a local database. An embedding model indexes every interaction, and the library exposes add, query and delete operations. The library is free, but you supply the embedding model. Hosted embedding APIs typically cost a few cents per thousand tokens, and a local model takes GPU time. He calls those fees a hidden expense for smaller teams that compounds over months.

The index files can be inspected with standard tools. The library enforces no access controls, though, so any process with file system access can read the memories, and permissions are left to the OS. In mem0’s configuration docs, the backend is a vector_store key naming a provider. The options include chroma, pgvector, qdrant, milvus, upstash_vector, azure_ai_search and vertex_ai_vector_search, with parameters such as embedding_model_dims, host, port and api_key. The author recommends mem0 for pipelines that need semantic similarity lookups across thousands of stored facts, run by teams with the GPU or API budget to pay for them.

LoreConvo puts all session data in one SQLite file in a directory you choose, with no cloud component. You can move the file, back it up or keep it under version control with the tools you already use. An MCP server exposes it to Claude Code, OpenAI Codex, Cursor and Hermes Agent. A .mcp.json in the project root is picked up automatically, with no per-client setup.

Recall runs on FTS5, SQLite’s built-in full-text index, so no embedding layer is needed. Session linking, project tagging and an auto-load hook carry context from one run to the next. For scripts that should not register an MCP tool, a Python fallback, save_to_loreconvo.py, reads and writes memories directly. The author pitches skill history tracking, tagging and linking to data engineers as a lightweight audit trail. It records which decisions shaped a pipeline, which schemas changed and why, and where a stalled refactor left off.

Saving is automatic. A hook runs at the end of every session and extracts a heuristic summary of decisions, tech-stack facts and open questions when the session has enough signal. The author calls that summary “best-effort, not a guarantee on every save.”

The free tier stops at fifty sessions. Pro costs $8 a month and removes the limit. It also adds semantic search, related-session discovery and team memory sharing. All of it runs against the same local file, with no per-call fees. The author says Pro costs a solo developer less than the extra API calls the cloud-native options typically generate. Team sharing works by exporting selected sessions to JSON for a teammate to import, with no central server involved.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.