Which memory provider should Hermes use?

An LLM pass extracts facts from each conversation, deduplicates them and stores them for semantic search, with optional reranking. The Hermes plugin runs against Mem0's cloud platform, a Mem0 server you host, or fully in-process with your own LLM and vector store. Since 27 October 2026 Mem0 sells no self-serve paid plan: the free plan covers most solo use if you bring your own LLM key, while graph memory, user profiles and SSO need a custom contract.

Retains full conversation turns, resolves entities and relationships into a knowledge graph, and recalls with several strategies at once (semantic, keyword, entity and time). Its reflect tool reasons across everything stored. Runs fully local with an embedded Postgres and any LLM, including open-weight ones, or on Hindsight Cloud with no monthly fee.

A context database from ByteDance's Volcengine. Memories, documents and skills live in a virtual filesystem (viking://) and load in tiers: a short abstract, then an overview, then full content only when it's needed. That's where the token savings come from. In Volcengine's May 2026 test, adding it to Hermes raised LoCoMo accuracy from 33.4% to 82.9% and cut input tokens by 34% and query time by 66%.

An LLM extracts entities and relations into a knowledge graph alongside a vector store, and recall walks both. The Hermes plugin adds remember, recall, forget, recap and code search, and shares memory with Cognee's Claude Code and Codex plugins. Local mode installs about 1 GB of dependencies.

Builds a memory graph per user (a profile plus a fact hierarchy) and pairs it with SuperRAG document search and connectors. The Hermes plugin captures each turn, recalls before each turn and re-injects profile facts periodically. The core engine is closed source; a free local server handles up to 10,000 documents.

Local memory built for coding agents. It captures sessions automatically and recalls with BM25 keyword search, optional on-device embeddings and a small knowledge graph, with no account or API key. One local server shares memory across 26 coding agents; the Hermes plugin connects to it and adds save, recall and search tools.

Turns chat into structured facts and summaries with project and session attribution, so recall stays short: about 721 tokens per answer in its own LoCoMo test. The Hermes provider uses Memori Cloud only and installs through Memori's own hermes-memori tool rather than the plugin catalog.

Models people, not just facts. Honcho stores messages, reasons about each user in the background, and answers questions about them on demand. Each Hermes profile becomes its own AI peer watching the same user, so a coding profile and a writing profile build separate views of you. Cloud or self-hosted.

From the AgentScope project. Memory is plain Markdown in a workspace folder (daily notes and digests) that you can read and edit, with BM25 recall that needs no embedding model. The Hermes provider recalls before each turn and records after it, in-process or through a local ReMe service. It adds no tools for the model.

A local SQLite memory layer with vector plus full-text search, episodic consolidation and a temporal knowledge graph, with no cloud at all. It exposes 41 tools to the model, the most of any provider here, covering personas, graph queries, import and export, and diagnostics.

A local SQLite fact store with full-text search, trust scores that learn from your feedback, and Holographic Reduced Representations for algebraic queries: everything about one entity, facts that link several, and contradiction checks. No API keys, no network calls and no LLM required. It used to ship inside Hermes and now installs from the catalog with the same data.