Exabase M-1 Memory Plugin for Hermes Agent

Exabase M-1 memory-provider integration for Hermes Agent.

About

Exabase Memory (M-1) is a self-organising memory engine for AI agents. It stores facts, preferences, and events, builds a living knowledge graph, resolves contradictions, and evolves with every interaction.

M-1 is SOTA on the leading AI memory benchmark (LongMemEval), with the highest recorded QA score, and using a small model. Read the research paper here.

System	Model	Score
M-1 (Exabase)	Gemini 3 Flash	96.4%
Mem0	Gemini 3 Pro	94.8%
Honcho	Gemini 3 Pro	92.6%
HydraDB	Gemini 3 Pro	90.79%
Supermemory	Gemini 3 Pro	85.2%

Exabase Memory powers memory in production apps like Fabric, used by 300,000+ people.

What it does

This plugin adds support for Exabase M-1 long-term memory provider to Hermes Agent. It allows agents to store conversation turns and inferred memories in Exabase, and provide context for future interactions.

Installation

Install the plugin from this repository and enable it in your Hermes Agent configuration:

hermes plugins install git@github.com:futurebrowser/hermes-exabase-plugin.git
hermes memory setup  # choose 'exabase'

Choose exabase as the memory provider when prompted and enter your Exabase API key.

Configuration

The plugin can be configured with optional parameters that control how memories are stored and retrieved.

• Base ID: The Exabase Base used for storing memories. This lets you scope memories and separate them by project, agent, or any other criteria. If not provided, no scoping will be used and memories will be stored in the default base.

• Precision: Controls the precision of memory retrieval. Higher precision means only the most relevant memories will be retrieved, while lower precision allows for more memories.

• Query expansion: Enabling query expansion allows the plugin to expand search queries with related terms, which broadens the search space and may help find more relevant memories at the cost of speed and Exabase credits.

• Result reranking: Uses an additional round of processing to rerank retrieved memories based on relevance to the query. This may improve the ranking of relevant memories at the cost of speed and Exabase credits.

Tools

• exabase_search searches long-term memories.
• exabase_remember stores a memory.

Completed conversation turns are sent to Exabase in the background for inferred memory extraction.