From e625d053a459c5b40262c070a959f6f8ad7905f9 Mon Sep 17 00:00:00 2001 From: roli-lpci Date: Tue, 15 Sep 2026 16:29:51 -0400 Subject: [PATCH 1/2] docs: add Fidelis Memory MCP integration --- integrations/fidelis.md | 56 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 56 insertions(+) create mode 100644 integrations/fidelis.md diff --git a/integrations/fidelis.md b/integrations/fidelis.md new file mode 100644 index 00000000..e1af220f --- /dev/null +++ b/integrations/fidelis.md @@ -0,0 +1,56 @@ +--- +layout: integration +name: Fidelis Memory +description: Local-first, zero-LLM agent memory for Haystack agents — BM25, dense-vector, and reciprocal-rank-fusion retrieval that returns original passages verbatim through a Model Context Protocol server. +authors: + - name: Hermes Labs + socials: + github: hermes-labs-ai +pypi: https://pypi.org/project/fidelis-memory/ +repo: https://github.com/hermes-labs-ai/fidelis +type: Custom Component +report_issue: https://github.com/hermes-labs-ai/fidelis/issues +version: Haystack 2.0 +toc: true +mcp: true +--- +### **Table of Contents** +- [Overview](#overview) +- [Installation](#installation) +- [Usage](#usage) +- [License](#license) + +## Overview + +[Fidelis](https://github.com/hermes-labs-ai/fidelis) is a local-first memory and retrieval service for AI agents. It stores notes and session context locally (`~/.cogito/`) and retrieves them with BM25, dense-vector, and reciprocal-rank-fusion (RRF) scoring, returning the original stored passages verbatim rather than paraphrasing them. The default retrieval path makes no LLM call. + +Fidelis ships as an MCP server (`fidelis mcp serve`, stdio transport) and is already published on the [official MCP Registry](https://registry.modelcontextprotocol.io/v0.1/servers/io.github.hermes-labs-ai%2Ffidelis-memory/versions/0.1.0) as `io.github.hermes-labs-ai/fidelis-memory`. A Haystack agent can connect to it the same way it connects to any other MCP server, using Haystack's own `mcp-haystack` integration (`MCPTool`/`MCPToolset` with `StdioServerInfo`), to give the agent a private, verbatim-recall memory backend instead of a hosted memory platform. + +On a checked-in 470-question LongMemEval-S retrieval run, Fidelis measured 83.2% R@1. + +## Installation + +```bash +pip install fidelis-memory mcp-haystack +``` + +## Usage + +```python +from haystack_integrations.tools.mcp import MCPTool, StdioServerInfo + +server_info = StdioServerInfo( + command="uvx", + args=["--from", "fidelis-memory", "fidelis", "mcp", "serve"], +) +tool = MCPTool(name="fidelis_memory", server_info=server_info) + +# Use directly, or add `tool` to a Haystack Agent's tools list +result = tool.invoke(query="what did we decide about the retrieval backend?") +``` + +See the [Fidelis README](https://github.com/hermes-labs-ai/fidelis#readme) for the full MCP tool surface and the [user-fit matrix](https://github.com/hermes-labs-ai/fidelis/blob/main/docs/user-fit.md) for supported workflows and prerequisites. + +## License + +MIT — see the [Fidelis repository](https://github.com/hermes-labs-ai/fidelis/blob/main/LICENSE). From ff8c6a4177dc2f6e2734abacaf8ec8780bfd318f Mon Sep 17 00:00:00 2001 From: roli-lpci Date: Thu, 17 Sep 2026 16:01:42 -0400 Subject: [PATCH 2/2] docs: fix Fidelis MCP setup and tool name --- integrations/fidelis.md | 18 +++++++++++------- 1 file changed, 11 insertions(+), 7 deletions(-) diff --git a/integrations/fidelis.md b/integrations/fidelis.md index e1af220f..9b65f52f 100644 --- a/integrations/fidelis.md +++ b/integrations/fidelis.md @@ -1,7 +1,7 @@ --- layout: integration name: Fidelis Memory -description: Local-first, zero-LLM agent memory for Haystack agents — BM25, dense-vector, and reciprocal-rank-fusion retrieval that returns original passages verbatim through a Model Context Protocol server. +description: Local-first, zero-LLM agent memory for Haystack agents, with dense-vector and optional BM25 plus reciprocal-rank-fusion retrieval through a Model Context Protocol server. authors: - name: Hermes Labs socials: @@ -22,28 +22,32 @@ mcp: true ## Overview -[Fidelis](https://github.com/hermes-labs-ai/fidelis) is a local-first memory and retrieval service for AI agents. It stores notes and session context locally (`~/.cogito/`) and retrieves them with BM25, dense-vector, and reciprocal-rank-fusion (RRF) scoring, returning the original stored passages verbatim rather than paraphrasing them. The default retrieval path makes no LLM call. +[Fidelis](https://github.com/hermes-labs-ai/fidelis) is a local-first memory and retrieval service for AI agents. It stores notes and session context locally (`~/.cogito/`) and supports dense-vector retrieval plus optional BM25 and reciprocal-rank-fusion (RRF) scoring. The default retrieval path makes no LLM call. Its MCP recall and query tools return evidence excerpts from the stored passages rather than generated summaries. -Fidelis ships as an MCP server (`fidelis mcp serve`, stdio transport) and is already published on the [official MCP Registry](https://registry.modelcontextprotocol.io/v0.1/servers/io.github.hermes-labs-ai%2Ffidelis-memory/versions/0.1.0) as `io.github.hermes-labs-ai/fidelis-memory`. A Haystack agent can connect to it the same way it connects to any other MCP server, using Haystack's own `mcp-haystack` integration (`MCPTool`/`MCPToolset` with `StdioServerInfo`), to give the agent a private, verbatim-recall memory backend instead of a hosted memory platform. +Fidelis ships as an MCP server (`fidelis mcp serve`, stdio transport) and is already published on the [official MCP Registry](https://registry.modelcontextprotocol.io/v0.1/servers/io.github.hermes-labs-ai%2Ffidelis-memory/versions/0.1.0) as `io.github.hermes-labs-ai/fidelis-memory`. A Haystack agent can connect to it the same way it connects to any other MCP server, using Haystack's own `mcp-haystack` integration (`MCPTool`/`MCPToolset` with `StdioServerInfo`), to give the agent a private, locally stored memory backend instead of a hosted memory platform. On a checked-in 470-question LongMemEval-S retrieval run, Fidelis measured 83.2% R@1. ## Installation ```bash -pip install fidelis-memory mcp-haystack +pip install "fidelis-memory[hybrid]==0.1.0" mcp-haystack +fidelis init +fidelis health ``` +`fidelis init` configures and starts the local `fidelis-server` used by the MCP bridge. Follow the [Fidelis quickstart](https://github.com/hermes-labs-ai/fidelis#quickstart) first to install its local Ollama embedding prerequisite. + ## Usage ```python from haystack_integrations.tools.mcp import MCPTool, StdioServerInfo server_info = StdioServerInfo( - command="uvx", - args=["--from", "fidelis-memory", "fidelis", "mcp", "serve"], + command="fidelis", + args=["mcp", "serve"], ) -tool = MCPTool(name="fidelis_memory", server_info=server_info) +tool = MCPTool(name="fidelis_recall", server_info=server_info) # Use directly, or add `tool` to a Haystack Agent's tools list result = tool.invoke(query="what did we decide about the retrieval backend?")