tam — total-agent-memory for Hermes

Long-term memory for Hermes Agent backed by total-agent-memory (TAM), over MCP.

Saves each completed turn (user message + final reply; no tool output) to TAM, skipping trivial prompts and very short exchanges. TAM's quality gate, when enabled, has the final say.
Before each turn, recalls relevant memories from TAM within a token budget.
Tools: tam_recall (search) and tam_save (store a durable fact, decision, solution, lesson or convention).
Shares one store with other agents on the same TAM (Claude Code, Codex CLI, Cursor, ...).

Local-first. By default Hermes starts your installed tam binary (MCP stdio), so data stays on your machine. The spawned process does not receive Hermes' provider API keys. Remote mode connects to a TAM HTTP endpoint or a TAM team server (URL + optional TAM_API_TOKEN) for shared team memory.

Quick start

pipx install total-agent-memory        # provides the `tam` command
hermes plugins install tam
hermes memory setup tam                # local/remote, command, data dir, project; tests the connection
hermes memory status

The plugin has no Python dependencies and never crashes Hermes: if TAM is missing or slow, the recall for that turn is empty, saves are queued with retry, and shutdown waits a bounded time.

Configuration reference, privacy details and the team-server walkthrough: https://github.com/vbcherepanov/hermes-tam-memory#readme

tam

What it adds

Tools 2

README

tam — total-agent-memory for Hermes

Quick start