MemoHood

What MemoHood does

Before each reply, MemoHood runs a hybrid search that combines FTS5 full-text search with Russian stemming and BGE-M3 vector search served through Cloudflare Workers AI, merges the lists with reciprocal rank fusion, and optionally reranks with Cohere when a key is set. After each turn, a background thread captures facts: explicit signals such as corrections, decisions and preferences are caught by free keyword scoring, and only ambiguous cases go to a cheap Gemini flash-lite call.

When a new fact contradicts an old one, a supersede classifier marks the old fact as replaced and keeps it in dated history that memohood_fetch can retrieve. Pinned facts such as names, allergies or an explicit remember-forever request are excluded from decay. A nightly hermes cron job applies Ebbinghaus-style confidence decay, deduplication and day-to-week-to-month rollups, and a spend ledger skips a step when a provider reaches its monthly cap. Version 1.1 adds an optional recall gate that is off by default, graph rerank and MMR-based post-recall.

Key features

• Automatic recall of relevant facts before every reply
• Hybrid FTS5 and BGE-M3 vector search with reciprocal rank fusion
• Free keyword capture of corrections, decisions and preferences, with an LLM only for unclear cases
• Supersede versioning that keeps replaced facts in dated history
• Pinned facts excluded from nightly confidence decay
• A spend ledger with monthly caps per provider

When to use it

• Keeping user preferences and decisions available across separate Hermes sessions
• Replacing an outdated fact when a user changes their mind, while keeping the history
• Running agent memory locally on SQLite with Russian-language morphology support

Who it is for: Hermes Agent users who want automatic, local long-term conversation memory, including for Russian-language dialogue.

How it fits with Hermes Agent

It is built for Hermes Agent as a memory provider selected through memory.provider, targeting hermes-agent 0.18 and later.

Requirements: hermes-agent 0.18 or newer; vector search uses BGE-M3 through Cloudflare Workers AI and optional reranking uses a Cohere key.

FAQ