Summary

Hermes has memory built in and it works — the memory tool saves facts to ~/.hermes/ and session_search gives cross-session recall. For basic use that's fine.

But the backend is flat files with no retrieval intelligence. As memory grows, you get:

• No semantic search — recall is keyword/FTS only, misses conceptually related memories
• No fact extraction — raw text is stored as-is, no structure or deduplication
• No entity resolution — "the user" and "Ben" are unrelated entries
• No relevance ranking — older or less relevant memories surface equally

What better local memory looks like

A richer local backend would store structured facts, resolve entities, and rank results by relevance — all running locally, no cloud dependency.

hindsight-hermes is a working example of this via the plugin system. It runs fully locally via Docker and gives Hermes structured fact extraction, entity resolution, and multi-strategy retrieval (semantic + BM25 + graph). PR #1811 is the integration.

The ask here is for this kind of capability to be accessible without needing an external plugin — either as a first-class memory backend option or by making the current backend more pluggable so alternatives are easier to drop in.

Related

• Feature: Structured Memory System — Typed Nodes, Graph Edges, and Hybrid Search #346 Structured Memory System
• Feature: Memory Storage Migration — Flat Files to SQLite with Scope, Importance & Timestamps #674 Memory Storage Migration
• Feature: Semantic Memory Recall — Vector Search with Composite Scoring (Similarity + Recency + Importance) #677 Semantic Memory Recall