Mnemosyne
3,301 starsCommunity-maintained local-first memory layer listed in Hermes' official catalog (human-reviewed, commit-pinned): SQLite with vector + FTS5 hybrid recall, episodic consolidation and a temporal knowledge graph — officially described as Zero cloud. The catalog declares 32 tools (mnemosyne_recall/remember/forget/graph_query/scratchpad_*/shared_*, etc.).
- First published：date not on record（Hermes 插件目錄只記載收錄時間與釘住的 commit SHA，未記載插件首次發佈日期，故不填）
- data as of：
- 2026-09-30
Host
Reads and writes a local SQLite memory store (zero cloud); but the catalog disclosure spells out three things: the mnemosyne_remember_media tool refuses local file paths unless MNEMOSYNE_MEDIA_ALLOWED_PATHS is set, and sends media only to an operator-configured OpenAI-compatible endpoint when modality_enabled is on; video uses ffmpeg/ffprobe from PATH; the embedding model (BAAI/bge-small-en-v1.5) downloads from Hugging Face on first use.
The plugin lives in the mnemosyne repo's integrations/hermes-catalog subdirectory, installed at a catalog-pinned commit, registering tools directly into Hermes' tool surface. The mnemosyne core is a standalone project, but this Hermes integration layer (tool registration, scratchpad and per-agent behavior) binds to the Hermes plugin API.
Risk
"Local-first" is not "never leaves": once the media modality is on, files within the allowed paths are sent to your configured external endpoint — a loose allowlist becomes an exfiltration channel. Memory-poisoning risk matches other memory layers (false memories written to SQLite get recalled repeatedly). The HF model download is a supply-chain node.
Install prompt
在 Hermes Agent 的插件目錄中以名稱安裝 mnemosyne（要求 Hermes 0.21.4 以上）；詳細設定見 mnemosyne-oss/mnemosyne 倉庫的官方說明。Paste this to your agent. It deliberately contains no tool subcommands — check the tool's official docs for commands.