mukundakatta/hermes-agentmemory

This project is a memory plugin for Hermes Agent, an open-source AI agent. The plugin lets the agent store and recall things from past conversations, but it does so in a way the author calls pull-model and auditable. The agent only retrieves memory when it explicitly asks for it, and every retrieval is written to a log file the user can read. The README spends most of its space explaining why this exists when Hermes already has memory backends like Mem0 and Honcho. The author argues those backends consolidate events into derived summaries in the background, which makes two things awkward. Deletes are not always real, because the original event can still influence a summary that was already built. And memory injection is opaque, because the user does not get to see what was pulled in until something goes wrong. This plugin, in contrast, never does background work, performs writes synchronously, deletes events fully, and writes a trace record every time it prefetches. Installation is a clone into the Hermes user-plugins directory, a pip install of the anthropic Python client, and a config command that points Hermes at the new provider. The summarizer calls the Claude API, so an ANTHROPIC_API_KEY is required. Other settings include the Claude model id, how many past events to pull per prefetch, the token budget for the summary, and the path of the trace log file. The plugin exposes three tools to the agent: recall to surface matching past events, forget to delete a session or a single event with no tombstone left behind, and drift to report on retrieval quality over a rolling window. A tail of the trace.jsonl file shows the intent, the event ids used, the summary, and the drift snapshot for each call. The README is honest about a trade-off: the first turn of each session pays a small delay because the summary is built on demand rather than pre-warmed. The plugin is released under the MIT license.