Fidelis Memory

Local-first, zero-LLM memory for Codex, Claude Code, and AI agents.

73.0% end-to-end QA on LongMemEval-S. 83.2% R@1 retrieval. $0/query. No LLM in the default retrieval path.

Stop re-explaining context to your agent. fidelis returns your original notes verbatim, local-first, fast, about 60 seconds to install. Your agent already calls an LLM to think; it should not need another one just to remember. Designed for developers. The default zero-LLM retrieval path does not send memory content to an LLM. The documented fidelis init service configuration also disables mem0 and Chroma telemetry. That can reduce third-party data exposure, but deployments still own their security and compliance assessment.

your notes / sessions
       ↓
local memory store      (~/.cogito/, fully local)
       ↓
fidelis retrieval       (BM25 + dense + RRF, no LLM)
       ↓
original passages       (verbatim, never rephrased)
       ↓
Codex / Claude Code / your agent

What fidelis is:

• fast - ~216 ms local retrieval (full benchmark mean; vector-only path is faster)
• cheap - $0/query retrieval cost
• private - local memory store by default
• faithful - original stored passages returned, not paraphrases
• proven - benchmarked on LongMemEval-S (470 questions, public benchmark), with raw evidence in experiments/zeroLLM-FLAGSHIP-evidence/
• installable - Codex or Claude Code via MCP in about 60 seconds

Quickstart

# 0. one-time: Ollama + the local embedder (~280 MB)
brew install ollama && ollama serve &
ollama pull nomic-embed-text

# 1. install Fidelis Memory from PyPI
python3 -m pip install "fidelis-memory==0.0.95"
fidelis init                  # background service (launchd / systemd)
fidelis watch ~/notes         # auto-ingests markdown
fidelis mcp install --client codex   # or omit for Claude Code
fidelis mcp serve             # runs the MCP server over stdio
# Restart your agent client. Memory is on.

Package-name note: install Hermes Labs' package as fidelis-memory. The import name and CLI remain fidelis. The separate PyPI project named fidelis belongs to NGdust/fidelis.

Linux users swap brew install ollama for the equivalent install from ollama.com. See Requirements.

Fidelis Memory 0.0.95 is also published in the official MCP Registry as io.github.hermes-labs-ai/fidelis-memory. Registry-aware clients can launch the same released server directly from PyPI:

uvx --from "fidelis-memory==0.0.95" fidelis mcp serve

This starts the MCP stdio process; run fidelis init first when the local Fidelis service and store have not already been configured. Version 0.0.94 introduced supported Codex MCP installation and context-sensitive orientation; 0.0.95 added the independently discoverable registry release.

What you notice immediately

After the four commands above, the next time you open Codex or Claude Code:

• It stops asking you to repeat context you already wrote down.
• You can ask "what did we decide last week about auth?" - and the answer cites your actual decision, not a generic OAuth lecture.
• Architecture rationale you wrote in a markdown file two months ago surfaces when relevant.
• Your project context carries across sessions instead of resetting at every new conversation.
• Failed migration notes, naming conventions, founder voice memos - all queryable in your agent's normal flow.

Most of fidelis's value is not the benchmark; it's not having to explain the same thing twice.

Most AI memory systems rewrite your notes

Most memory systems rephrase content on the way out. The specific fact gets summarized into something general. fidelis solves this structurally - there is no LLM in the default retrieval path, so the store returns exactly what you put in.

You store:

auth tokens expire after 3600 seconds.
The 3600s window is non-configurable in our current contract.

A lossy memory layer may return:

authentication has a configurable timeout

fidelis returns:

auth tokens expire after 3600 seconds.
The 3600s window is non-configurable in our current contract.

The non-configurable qualifier survives. So does every other detail you wrote down.

What this enables in Codex, Claude Code, and GitHub Copilot CLI

Once fidelis mcp install --client codex, --client copilot, or the default Claude install is run, ask your agent:

• "What did we decide about auth?"
• "What failed last time we tried this migration?"
• "Which billing constraint was non-configurable?"
• "What did I say about Sarah's onboarding flow?"

The MCP fidelis_recall tool gives the agent the original passages before it composes an answer, not paraphrased summaries. The answer can stay grounded in what you wrote, with the qualifiers intact.

fidelis retrieves memory without an LLM. Your agent still uses its normal LLM to answer using the retrieved context. "Zero-LLM" applies to the memory hot path, not to your agent.

GitHub Copilot CLI

Copilot CLI loads MCP servers from mcp-config.json in its configuration directory (~/.copilot by default, or $COPILOT_HOME). Fidelis writes the documented stdio entry there atomically, backing up any existing file and leaving other servers untouched:

Unreleased. --client copilot is on main and not in the pinned 0.0.95 package installed in the Quickstart; it ships in the next release. Install from source to use it today.

fidelis mcp install --client copilot     # writes ~/.copilot/mcp-config.json
copilot                                  # restart, then /mcp list shows "fidelis"
                                         # /mcp show fidelis lists its tools
fidelis mcp uninstall --client copilot   # removes only the fidelis entry

Use --settings /path/to/mcp-config.json to target a different file. The copilot binary is not required at install time; if you prefer the host CLI, the equivalent registration is copilot mcp add fidelis -- "$(python3 -c 'import sys;print(sys.executable)')" "$(python3 -c 'import fidelis.mcp_cmd as m;print(m.MCP_SERVER_FILE)')". Copilot does not currently expose a hook or automatic-recall mechanism to third-party servers, so recall happens when the agent calls the fidelis_recall, fidelis_orient, or fidelis_health tools.

Use cases & ROI

Three concrete reasons teams pick fidelis over hosted memory:

• Cost reduction. Stop paying for redundant context-window tokens on every turn. Memory lives on disk; the agent pulls only what's relevant per query. At a few thousand calls/day the math against per-query memory APIs adds up fast.
• Local data boundary. The default zero-LLM path keeps notes and retrieval on the local machine, reducing third-party processor exposure. This architecture does not by itself confer SOC 2 or HIPAA compliance.
• Team context. Agents that remember historical decisions, naming conventions, failed migrations, and the qualifiers on those decisions. The non-configurable detail you wrote down two months ago surfaces when relevant, in the founder's voice, not paraphrased.

How it fits

The diagram is at the top. Codex and Claude Code are the fastest paths to value. The retrieval engine is agent-agnostic - pair it with any LLM client. Codex registration uses its supported codex mcp CLI, and the resulting server configuration is shared by the Codex desktop app, CLI, and IDE extension on that host.

Benchmarks

LongMemEval-S, 470 questions, public benchmark.

For context: published Mem0 results on LongMemEval-S are in the ~66–70% end-to-end QA range; Zep is 71.2%; Supermemory is 81.6%; full GPT-4o on raw context (no memory system) is 60.2%. fidelis reaches 73.0% with no LLM in the default retrieval path.

Raw evidence: retrieval aggregate · end-to-end QA summary

The QA tier wraps your existing LLM with a 140–180-token system prompt - the Fidelis Scaffold. See docs/scaffold.md.

Verify the zero-LLM claim yourself

# Unset any LLM API keys for this shell
unset OPENAI_API_KEY ANTHROPIC_API_KEY DASHSCOPE_API_KEY

# Optional: drop your network. Ollama runs on 127.0.0.1:11434 (loopback).

# `recall-hybrid` is the explicit-tier command. zero_llm is the default.
fidelis recall-hybrid "what did the user say about Sarah" --tier zero_llm
tail ~/.fidelis/server.log

The default zero_llm tier never makes an outbound LLM call. Optional --tier filter and --tier flagship modes do call an LLM, but only to select integer pointers - the server dereferences those pointers to the original stored text. The LLM cannot rephrase memory content.

Context-sensitive orientation (MCP)