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)

The bundled MCP server also exposes fidelis_orient. It recognizes when a turn invokes prior work—even when it is a statement such as “I need to remember our Fidelis work”—and selects a bounded evidence lane for identity, maintenance, conceptual reuse, comparison, decisions, historical state, or current state. The returned orientation is a derived index; retrieved records remain verbatim evidence with their existing IDs and metadata. Unrelated turns explicitly abstain without calling the memory server.

Requirements

• macOS or Linux (Windows not yet supported)

• Python 3.10+

• Ollama running locally with nomic-embed-text pulled (~280 MB):

  bash

  Copy

  brew install ollama && ollama serve &
  ollama pull nomic-embed-text   # ~280 MB, one-time
  

The full init-to-first-recall cycle is under 60 seconds once Ollama is up. No memory API keys required.

Quick reference

fidelis recall "what did the user say about Sarah"
fidelis query  "Sarah" --limit 5
fidelis add    "raw text to extract into memories"
fidelis health
fidelis seed   ~/memory/   ~/notes/

fidelis add normally stores facts produced by the configured extraction model. If extraction returns no facts, Fidelis preserves the original input verbatim instead of silently losing it. The command still exits 0 because the write succeeded, but stdout reports a stable degraded status:

status=stored degraded=verbatim-fallback-empty-extraction id=<uuid> count=1

Automation that requires successful extraction must inspect degraded; exit 0 means the memory was stored, not necessarily that extraction succeeded. Because mem0 does not distinguish a swallowed extractor failure from a legitimate zero-fact result, the fallback intentionally favors durability.

Python helper for direct integration:

from fidelis.augment import augment
from anthropic import Anthropic

client = Anthropic()
answer = augment(
    question="What did I say about Sarah?",
    qtype="single-session-user",
    llm_call=lambda system, user: client.messages.create(
        model="claude-haiku-4-5",  # any current Claude Messages model works
        system=system,
        messages=[{"role": "user", "content": user}],
        max_tokens=512,
    ).content[0].text,
)

What's running on your machine

After fidelis init:

• Service: fidelis-server runs at http://127.0.0.1:19420 under your OS service manager (launchd on macOS, systemd on Linux). Auto-starts on boot. Logs at ~/.fidelis/server.log.
• Storage: Chroma + SQLite at ~/.cogito/ (the directory name is preserved from the project's pre-rename codename for v0.0.x compatibility - it will move to ~/.fidelis/ in a later major bump). No data leaves your machine in the default zero-LLM path.
• MCP: after installing for your selected client, Codex or Claude Code sees four tools: fidelis_recall, fidelis_query, fidelis_health, and fidelis_orient.

To stop: fidelis init --uninstall. To wipe: rm -rf ~/.cogito ~/.fidelis.

Known limitations (v0.0.95)

• Pre-release. Python function names and CLI commands may change. Pin the version if you build on it.
• Best on macOS Sequoia / Ubuntu 24.04 LTS. Other OSes likely work but aren't gate-tested.
• Direct server launches disable mem0 telemetry by default. This matches the service installed by fidelis init and avoids telemetry exit handlers delaying graceful shutdown. An explicit MEM0_TELEMETRY=True still opts in. For the same boundary across Chroma, set ANONYMIZED_TELEMETRY=False and CHROMA_TELEMETRY_DISABLED=True before a direct launch; fidelis init includes all three settings automatically.
• Temporal-reasoning and preference questions are the weakest qtypes in the QA scaffold (TR ~58%, Pref ~37% on the full eval). Single-session and knowledge-update qtypes are strong (95–100%).
• The optional LLM tier ("flagship" mode) currently escalates ~80% of queries instead of the intended ~10% - an 8× cost miss we're transparent about. The default zero-LLM tier is unaffected.
• qwen3.5:9b in thinking mode does not reliably follow the literal hedge instruction in the Fidelis Scaffold. Use Claude, an OpenAI-format API, or non-thinking-mode local models for reliable hedging.

What this turns into over time

Day 1: drop notes into ~/notes, run the four commands. Day 2: ask your agent about yesterday's decision - the answer cites your original passage. Day 7: your agent starts carrying project context across sessions; you stop re-explaining.

Useful for solo builders today; relevant for teams that need memory to stay local tomorrow.

Fidelis Memory for teams

fidelis is open-source under MIT and free for any use, including commercial. If your team has deployment requirements that the OSS path does not yet cover (centralized memory, multi-namespace isolation, custom authentication), write to founders@hermes-labs.ai.

For technical users

• docs/full-reference.md - full architecture, hybrid recall tiers, local server endpoints, troubleshooting
• docs/scaffold.md - Fidelis Scaffold contract + drift-detection markers
• experiments/zeroLLM-FLAGSHIP-evidence/ - raw eval JSONs + machine-readable SUMMARY (per-qtype breakdowns, Wilson CI, F1/F1B baselines)

License

MIT. Built by Hermes Labs (Roli Bosch). Issues + PRs welcome.

About Hermes Labs

Hermes Labs develops open-source reliability, evaluation, memory, and containment tools for AI agents. Fidelis is its local-first memory project. Other public software is listed at github.com/hermes-labs-ai, with research artifacts published separately on Zenodo.

For enterprise deployments and AI-reliability engagements: roli@hermes-labs.ai · hermes-labs.ai

On naming. Hermes Labs is named for Hermes, the Greek messenger god - patron of communication and interpretation, the herald who carries meaning between worlds. The thread to the work: hermeneutics, the theory of interpretation that takes its name from Hermes, is the philosophical anchor for an AI reliability engineering studio whose substrate is linguistic. Not affiliated with NousResearch's Hermes LLM line or their hermes-agent framework - different companies, different work.

Founder: Rolando (Roli) Bosch. Site: hermes-labs.ai Citation: Bosch, R. (2026). Hermes Labs: AI reliability infrastructure for autonomous agents. https://hermes-labs.ai

Quantitative source for the Fidelis claims above: the 470-question LongMemEval-S aggregate and Wilson interval in experiments/zeroLLM-FLAGSHIP-evidence/, evaluated 2026-04-24.