Documentation

What is TAM

total-agent-memory (TAM) gives your AI coding agent a memory that survives between sessions, projects and even different tools.

The problem

Every new chat with Claude Code, Codex CLI or Cursor starts from zero. The agent does not know that yesterday you picked PostgreSQL over MongoDB, that the staging deploy needs a VPN, or that a certain migration must never be reverted. You explain it again, every time.

What TAM does

TAM is a small server that your agent talks to through the Model Context Protocol (MCP), the standard way AI tools call external tools. Once connected, the agent can:

Save things worth keeping: decisions and why they were made, fixes, lessons, conventions, facts.
Recall them later by meaning, not just by exact words, in any session and any project.
Resume work: at the start of a session it gets a short summary of where you stopped, what comes next and what to avoid.

You mostly talk to your agent in plain language (“remember that…”, “what did we decide about…?”). The agent picks the right tool.

Day 1   You:    Remember: we use pgvector, not ChromaDB, because we need per-tenant row-level security.
        Agent:  memory_save(type="decision", …)  ✓

Day 4   You:    Why did we pick pgvector again?
        Agent:  memory_recall(query="vector database choice")
                → "Chose pgvector over ChromaDB. WHY: single Postgres, per-tenant RLS." (saved 4 days ago)

What makes it different

Local-first. Your memory is a SQLite file on your machine (~/.tam/memory.db). Search runs locally with small embedding models. No account, no telemetry.
No LLM needed to save or search. In the default fast mode, saving and searching make no calls to a language model and no network requests. An LLM is optional and only used for extra features you switch on.
Works with many agents. Claude Code, Codex CLI, Cursor, Cline, Continue, Windsurf, Gemini CLI, OpenCode and Claude Desktop; Aider through a small command-line bridge. They can all share one memory.
More than notes. A knowledge graph, facts with validity dates, error-to-rule learning, task workflows and a local dashboard at http://127.0.0.1:37737.
Team mode. Since 14.0 one server can hold personal, team and company-wide memory with per-person tokens and full edit history. See Team server.

How good is recall?

On the public LoCoMo and LongMemEval benchmarks, a head-to-head grading of TAM against Mem0 Platform on the same held-out questions found no statistically significant difference at the same answering model. That is not a claim of equivalence or of first place; the sample cannot rule out differences of a few points. The full report describes the protocol, judges and how to reproduce it.

Open source

TAM is MIT-licensed. The source, issues and releases live on GitHub.

Ready? Go to the Quick start.