Fidelis Memory 0.1.0

• Promote Fidelis Memory from the 0.0.x sequence to its first minor release.
  The supported product contract is local-first, verbatim memory retrieval for
  Codex, Claude Code, GitHub Copilot CLI, Gemini CLI, and OpenClaw on the
  currently gate-tested macOS and Ubuntu paths.
• Coordinate package, Python, citation, CodeMeta, MCP Registry, Gemini
  extension, installation-documentation, and public-install-test versions at
  0.1.0.
• Add a public user-fit matrix, a release-readiness record
  with explicit acceptance criteria, and an outcome-gated 0.2.0 roadmap.
• Replace the manually maintained passing-test-count badge with the live GitHub
  Actions status badge so release documentation cannot advertise an obsolete
  count.
• Preserve the historical v0.1.0 and v0.2.0 changelog entries under the
  clearly labeled cogito-ergo predecessor-package section so they cannot be
  mistaken for current Fidelis package releases.

Fidelis Memory v0.0.97

• Release the Gemini CLI extension: v0.0.96 was tagged before
  gemini-extension.json landed, so the bare-URL install resolved the latest
  GitHub release and failed with "Configuration file not found". v0.0.97 is the
  first tag and release archive that carry the manifest.

• Pin the user_id namespace boundary in tests and document that it is not an
  identity; add glama.json for the Glama registry ownership claim.

• Package Fidelis as a native Gemini CLI extension: gemini-extension.json
  at the repository root launches the MCP Registry package
  (uvx --from fidelis-memory==0.0.97 fidelis mcp serve) and GEMINI.md
  tells the model when to call fidelis_orient and fidelis_recall.
  gemini extensions install https://github.com/hermes-labs-ai/fidelis
  registers the server without a manual pip install.
  tests/test_gemini_extension.py binds the manifest to server.json and
  pyproject.toml so a release bump cannot leave it on an older wheel.

Fidelis Memory v0.0.95

What's Changed

• feat: publish Fidelis 0.0.95 through MCP Registry by @roli-lpci in #13

Full Changelog: v0.0.94...v0.0.95

Fidelis Memory v0.0.94

Fidelis Memory 0.0.94 makes the merged Codex integration available through the public package channel.

What is new

• Supported Codex MCP installation and removal through the shared codex mcp configuration.
• Exact server-ownership checks preserve unrelated MCP entries during install and uninstall.
• New fidelis_orient MCP tool selects a bounded evidence lane, returns retrieved records without rephrasing them, and abstains when a turn lacks a known referent or historical-context cue.
• Updated citation, CodeMeta, and agent-facing discovery metadata.

Install

python3 -m pip install "fidelis-memory==0.0.94"
fidelis init
fidelis mcp install --client codex

Omit --client codex to install for Claude Code.

The orientation packet is a derived index over retrieved evidence; it is not a synthesized source of truth. This release exposes Fidelis capabilities only and does not imply that separate plugins already invoke fidelis_orient.

PyPI: https://pypi.org/project/fidelis-memory/0.0.94/

Fidelis Memory v0.0.93

Fidelis Memory 0.0.93

This release publishes the Hermes Labs distribution to PyPI as fidelis-memory.

• Uses Fidelis Memory as the user-facing product name.
• Preserves the fidelis repository, Python import, and CLI names for compatibility.
• Makes the version-pinned PyPI installation path the primary install route.

Install the exact release:

python3 -m pip install "fidelis-memory==0.0.93"

PyPI: https://pypi.org/project/fidelis-memory/0.0.93/

Full changes: v0.0.92...v0.0.93

fidelis v0.0.92

A reliability release for the public, GitHub-source Fidelis package.

What changed

• No silent loss on empty extraction: /add stores the original input verbatim when the extraction model returns no facts, reports degraded=verbatim-fallback-empty-extraction, and applies the same rule during queue replay.
• Clean direct-server shutdown: direct fidelis-server launches now default mem0 telemetry off, matching the generated service configuration and preventing telemetry exit handlers from delaying SIGTERM. Explicit MEM0_TELEMETRY=True still opts in.
• Current Python coverage: CI now installs and tests Python 3.10 through 3.14 on both Ubuntu and macOS.
• Install truth: documentation consistently points to this tagged GitHub source and warns that the PyPI project named fidelis is unrelated.

Install

git clone --branch v0.0.92 --depth 1 https://github.com/hermes-labs-ai/fidelis.git
cd fidelis
python3 -m pip install .
fidelis init

Evidence boundary

The fallback evaluation uses synthetic fixtures to establish mechanism conformance. This release makes no new natural-corpus accuracy claim. Fidelis remains pre-release software; pin the tag when integrating it.

fidelis v0.0.91

Zero-LLM agent memory: BM25 + dense vectors + reciprocal rank fusion, verbatim recall, fully local. Benchmarked on LongMemEval-S.

v0.0.9 — zero-LLM agent memory for Claude Code: 73% LongMemEval-S, $0/query

v0.0.9 — zero-LLM agent memory for Claude Code: 73% LongMemEval-S, $0/query

fidelis is the agent memory layer that structurally cannot rephrase your notes. Local by default, $0 per query, returned verbatim. Built for Claude Code, Cursor, Cline, and any LLM agent that needs faithful memory of what you actually wrote down.

73.0% end-to-end QA on LongMemEval-S (470 questions, Wilson 95% CI [68.7%, 77.0%]) · 83.2% R@1 retrieval · $0/query · 486 tests · Hermetic CI on Ubuntu + macOS × Python 3.10/3.11/3.12 · MIT.

This is the flagship release of Hermes Labs — the first publicly shippable artifact of a Series-A-grade AI infrastructure stack built solo, AI-amplified, at the reliability bar a funded team would call v1.

Install (3 minutes)

brew install ollama && ollama serve &
ollama pull nomic-embed-text       # ~280 MB, one-time
pip install git+https://github.com/hermes-labs-ai/fidelis@v0.0.9
fidelis init
fidelis mcp install                # only if you use Claude Code
# Restart Claude Code. Memory is on.

Linux users swap brew install ollama for the install path at https://ollama.com.

Why fidelis exists

Most agent memory layers silently rephrase what you store. You write "auth tokens expire after 3600 seconds, non-configurable in our current contract" and the memory layer hands back "authentication has a configurable timeout" an hour later. The qualifier is the point. The rephrase is the loss — the original meaning is gone, the agent answers wrong, and you don't notice because the rephrase reads plausible.

fidelis is built around a single architectural commitment: integer-pointer fidelity. The optional filter LLM never sees free-text memory content. It sees a numbered candidate list and outputs integers. The server dereferences those integers to the verbatim stored passages. The LLM is structurally incapable of rephrasing, summarizing, or hallucinating into the returned text.

This is structural, not prompting. You don't ask the LLM nicely to preserve your wording. You make it impossible for the LLM to alter your wording.

The default retrieval tier removes the LLM from the hot path entirely — BM25 + dense + RRF on local Chroma, embeddings via local Ollama, $0 per query, ~216 ms mean latency. The optional filter and flagship tiers add LLM ranking when accuracy matters more than cost or latency, but the integer-pointer guarantee survives at every tier.

This is what faithful agent memory looks like: your notes leave your machine only if you explicitly send them, and they come back exactly as you wrote them down.

How fidelis compares

	fidelis	mem0	Zep	Supermemory
LongMemEval-S end-to-end QA	73.0%	~66–70%	71.2%	81.6%
Cost per query (retrieval)	$0	per-query (cloud)	per-query (cloud)	per-query (cloud)
Where data lives	local disk	their cloud	their cloud	their cloud
Returns original passages verbatim	yes	no (LLM-rephrased)	no	no
Open source	yes (MIT)	partial	no	no
Runs fully offline	yes (default)	no	no	no

For raw context: published Mem0 results on LongMemEval-S sit 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: bench/runs/zeroLLM-full-20260424/aggregate.json.

What's new in v0.0.9

This is the first publicly shippable release. The reliability surface is what makes it ship-ready, not just feature-ready.

• Atomic queue writes — tempfile + fsync + os.replace. SIGKILL mid-write cannot corrupt user data.
• Per-record dead-letter queue at MAX_ATTEMPTS=5 with forensic preservation. A poisoned record cannot keep the replay loop hot indefinitely.
• Exponential-backoff replay — 60 s base → 30 min cap, doubles on no-progress sweeps, resets on any successful replay.
• Idempotent retries — duplicate-id chroma inserts treated as success rather than another failure.
• Graceful shutdown — SIGTERM and SIGINT handlers call httpd.shutdown() cleanly so SQLite WAL is checkpointed before process exit. Live-fire test spawning a real fidelis-server and asserting the clean-stop marker is included.
• Ingestion backpressure — fidelis watch caps at 50-file batches with a /health pressure check and a 1 GB session byte cap, so dropping 10,000 PDFs into a watched directory cannot OOM the user's machine.
• Hermetic CI — Ubuntu + macOS × Python 3.10/3.11/3.12, six fresh-install proofs per push, plus an sdist + wheel build job.
• Migration safety — fidelis init preflights Ollama, boots out the legacy ai.hermeslabs.cogito-server plist, warns on stale mistral:7b config without auto-rewriting. Telemetry-kill envs in plist + systemd unit prevent the chromadb posthog file-descriptor leak that EMFILE-crashes restarted services.
• Workaround for upstream mem0 2.0.0 bug — mem0's Memory.search() returns identical score=1.0 for every result; fidelis bypasses the wrapper and calls vector_store.search() directly. Issue filed upstream.
• Tests — 486 collected; 481 + 1 xpassed locally in 12 minutes. LLM-backed integration tests run separately.

Previously developed under the internal codename cogito-ergo. The package, CLI, MCP server, and HTTP service are all named fidelis as of v0.0.9.

Verify the zero-LLM claim yourself

unset OPENAI_API_KEY ANTHROPIC_API_KEY DASHSCOPE_API_KEY
fidelis recall-hybrid "what did we decide about auth" --tier zero_llm

The default zero_llm tier never makes an outbound LLM call. Optional --tier filter and --tier flagship modes 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 at any tier.

Common questions

What is fidelis?
A local agent memory system that returns your original notes verbatim, with no LLM in the retrieval hot path. Built for developers and teams who need agent memory to stay on disk — for compliance reasons (SOC2 / HIPAA scoping), cost reasons ($0 per query), or latency reasons (no network round trip).

How is fidelis different from mem0, Zep, or Supermemory?
Those are hosted services that send your notes to their cloud and rephrase them through an LLM on the way out. fidelis runs locally and structurally cannot rephrase the stored text — the optional filter LLM tier outputs only integer pointers, and the server dereferences those pointers to verbatim text.

Does fidelis work fully offline?
Yes. The default zero-LLM retrieval tier never makes an outbound network call. Embeddings come from a local Ollama instance. Unset your API keys, pull the network plug, and fidelis recall-hybrid --tier zero_llm still works.

What's "integer-pointer fidelity"?
The optional filter LLM never sees free-text memory content. It sees a numbered candidate list and outputs integers. The server dereferences those integers to the verbatim stored passages. The LLM is structurally incapable of rephrasing, summarizing, or hallucinating into the returned text. This is structural fidelity, not prompting fidelity — a property of the architecture, not a property of the prompt.

Is fidelis production-ready?
v0.0.9 is the first publicly shippable release. Hermetic CI, atomic writes, dead-letter queue, graceful shutdown, and ingestion backpressure are landed. Not yet on PyPI; install via pip install git+https://github.com/hermes-labs-ai/fidelis@v0.0.9. Honest known-limitations below.

What's the LongMemEval-S score?
73.0% end-to-end QA accuracy with Wilson 95% CI [68.7%, 77.0%], on the full 470-question public benchmark, with the Fidelis Scaffold wrapping any current Anthropic Claude model. Raw evidence: bench/runs/zeroLLM-full-20260424/aggregate.json.

Who built fidelis?
fidelis was built by Rolando (Roli) Bosch, solo founder of Hermes Labs. Hermes Labs is a one-person AI infrastructure company shipping at a reliability bar a funded team would call v1. fidelis is the flagship release of the Hermes Labs agent-reliability stack alongside lintlang (static analysis for AI agent configurations), hermes-seal (cryptographic containment for autonomous agents), and the Hermes scaffold + rubric tooling.

Known limitations (v0.0.9, honest list)

• 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 are not gate-tested.
• Temporal-reasoning and preference qtypes are weakest in the QA scaffold (TR ~58%, Pref ~37% on the full eval).
• Optional flagship LLM tier escalates ~80% of queries instead of the intended ~10%, an 8× cost miss the project is transparent about. The default zero-LLM tier is unaffected.
• Non-technical install bar not met. Requires Ollama and a Python environment. Single-command bootstrap is post-1.0.

About

fidelis is built and maintained by Rolando (Roli) Bosch — solo founder of Hermes Labs, https://hermes-labs.ai.

Hermes Labs is a one-person AI infrastructure company. The thesis is that a solo founder, AI-amplified, can ship at a reliability bar that historically required a funded team of ten — atomic writes, dead-letter queues, hermetic CI, graceful shutdown, ingestion backpressure, honest known-limitations on every release. fidelis is the flagship expression of that thesis.

Other Hermes Labs OSS infrastructure: lintlang (static analysis for AI agent configurations), hermes-seal (cryptographic containment for autonomous agents), and the Hermes scaffold + rubric tooling for AI epistemic reliability.

If you're a developer building with fidelis, issues an...