FAQ

Do I need an API key or an account?

No. Saving and searching run locally. An LLM (local Ollama or a cloud provider) is optional and only powers extra features such as background enrichment and memory_answer.

Does it send my code or notes anywhere?

Not by default. See Privacy & local-first for the exact list of optional features that make network calls.

Which clients are supported?

Claude Code, Codex CLI, Cursor, Cline, Continue, Windsurf, Gemini CLI and OpenCode have an automatic connect command. Claude Desktop and any other MCP client work with a manual config entry. Aider has no MCP support; it can use the lookup-memory command. See Quick start.

Can several clients share one memory?

Yes. They all start the same server against the same ~/.tam/memory.db. Save in Claude Code, recall in Cursor.

Can my team share memory?

Yes, with the team server: one server, a personal token per person, and personal, team and company-wide (shared) memory with edit history.

How much disk and RAM does it use?

The base install is about 113 MB of Python wheels plus the embedding model (about 220 MB for the default multilingual model). The database grows with what you save, including vectors and search indexes for every record. The optional reranker adds PyTorch (about 3 GB on Linux). A team server with three warm workers needs at least 4 GiB of RAM.

Is it fast?

In fast mode, saving and searching involve no LLM calls and no network. The first call after a start loads the embedding model, so it is slower than the rest. Measured latencies, with the hardware and dataset they were taken on, are in the repository’s benchmark reports.

How does it compare with Mem0, Zep, Letta and others?

On LoCoMo and LongMemEval, a head-to-head grading of TAM and Mem0 Platform on the same held-out questions found no statistically significant difference at the same answering model. That does not show equivalence and is not a ranking. The report explains the protocol and its limits. The main practical difference: TAM retrieves locally and calls no LLM when it writes or searches.

Will it fill up with junk?

The agent saves what you and its instructions tell it to. Good habits help (Your first day). Operational records (recovery, auto-extract) are hidden from search by default, and memory_forget archives stale records that have never been recalled.

Can I edit or delete a memory?

Yes: ask the agent, or use memory_update and memory_delete. The dashboard’s knowledge browser shows what is stored.

What happens to my data when I upgrade?

Upgrades apply database migrations in place and keep your data. Take a backup anyway; see Upgrade & backup.

Is it free?

TAM is open source under the MIT licence.