hermes-memory-pgvector

Current version: 1.0.2.

Postgres + pgvector memory provider for hermes-agent. A shared memory substrate for a fleet of cooperating hermes-agent minions — built on Postgres and a single embedding endpoint you probably already run, with no LLM in the memory hot path.

each minion → X-Hermes-Session-Key: <theme>
            → hermes-agent gateway
            → pgvector plugin
                ├── memory_entries  (mirrors built-in MEMORY.md / USER.md per theme)
                └── conversations   (every substantive turn, semantically searchable)

Why it exists

Existing memory providers each solve a piece of the problem; the gap for fleet deployments is wide:

• Built-in memory tool persists to per-host MEMORY.md / USER.md. Two minions on the same host stomp on each other; minions on different hosts have no shared substrate.
• Honcho offers cross-session user modelling but requires a full external service, an LLM in the memory hot path for its deriver + dialectic loops, and its own ontology layered on top of the built-in tool. In high-concurrency fleet use it produces…