Howdex is the open verification layer for agent know-how. It turns the expensive, hard-won experience of your best AI agents into verified agent procedures that any agent can reuse, any team can audit, and any model, frontier or open, can execute.
One expensive discovery. Infinite cheap executions.
Howdex turns execution traces into portable, receipt-backed procedures that agents can reuse and enterprises can audit.
Today your agents are amnesiacs. An agent solves a gnarly problem — recovers a broken service, completes a refund-and-rebooking flow, untangles a build — and the moment the session ends, that competence evaporates. Tomorrow it starts from zero: re-reading the same files, repeating the same dead-ends, burning the same tokens. You are paying frontier prices, over and over, to re-learn things you already knew.
Howdex fixes that. It watches your agents succeed, distills the repeatable procedure behind the success, proves it works with attached receipts, and hands it back as clean operational guidance the next agent can follow, deterministically, locally, and portably across every model and framework you use.
We don't ask you to take this on faith. On a real Docker Compose recovery task, a fresh agent must read five files, infer a cross-file configuration dependency, and repair the right one to bring a service healthy:
| Arm | Success rate | Avg attempts |
|---|---|---|
| Fresh agent, no memory | 35% (7/20) | 13.00 |
| Same agent + Howdex | 90% (18/20) | 8.75 |
- n = 20 per arm, byte-identical task framing across both arms (verified by prompt hash).
source_pasted: 0/20— the agent reconstructed the fix from learned operational facts. It was never handed the answer. This is real transfer of know-how, not a leaked solution.- Reproduce it yourself:
make bench-docker-n20
A 55-point jump in success and a 33% drop in attempts, from memory alone, on a model that had never seen the task.
pip install howdex-ai
# or, straight from source:
pip install git+https://github.com/rossbuckley1990-hash/Howdex.gitfrom howdex import Howdex
mem = Howdex() # local SQLite at ~/.howdex — nothing leaves your machine
# 1. Capture a successful run (wrap the tool calls your agent already makes)
mem.start_session("recover broken compose service")
mem.log_tool_call("read_file", {"path": "compose.yml"}, observation="ok")
mem.log_tool_call("edit_file", {"path": "runtime.env"}, observation="set HEALTH_MODE=ready")
mem.log_tool_call("run", {"cmd": "docker compose up -d --build"}, observation="healthy")
mem.end_session("success")
# 2. Distill the reusable procedure
mem.learn()
# 3. On a new, related task, get clean, inspectable guidance
print(mem.guidance("service /health returns 503 after deploy"))Plug it into any agent via MCP — no code at all:
howdex mcp --db ~/.howdex/howdex.db --codex ./codexPoint Claude Desktop, Cursor, Windsurf, or any MCP client at it, and your agent gains remember, learn, and guidance as first-class tools.
Most "agent memory" stores facts (who the user is, what they like) by asking an LLM to summarize conversations, opaque, non-deterministic, and locked to one cloud. Howdex stores how the work was actually done — and proves it.
| Conversational / factual memory (Mem0, Zep, Letta) | Howdex | |
|---|---|---|
| What it stores | facts, preferences, chat history | reusable procedures — how to do the work |
| How it's built | LLM summarization | deterministic — no LLM in the loop |
| Inspectable & auditable | rarely | yes — full provenance + receipts |
| Runs locally, no model required | varies | yes — numpy-only core |
| Verifies before trusting | no | yes — the Codex |
| Portable across models & clouds | varies | yes — you own the loop |
Howdex isn't a competitor to factual memory, it's the procedural layer that sits alongside it and does the thing none of them do: make competence repeatable, portable, and provable.
Drop-in capture around the tool calls you already make. A 5-minute local setup, an MCP server, and one dependency (numpy). No cloud account, no vector database to operate, no API key to babysit. Your agent gets faster and cheaper on every task family you touch, and you can read exactly what it learned, because nothing is a black box.
One integration, every stack. Howdex ships adapters for LangChain, LangGraph, CrewAI, AutoGen, and the OpenAI Agents SDK, plus a generic decorator for anything else and a full MCP server. Add procedural memory to your framework of choice without rewriting your agent.
Howdex is local-first, vendor-neutral, and audit-friendly by design. Procedures stay inside your perimeter. Every verified procedure carries provenance, a verifier, and a receipt — so when an agent recalls "how we do X," you can prove where that came from and that it actually worked. Built-in governance (codex lint, policy-check, signed attestations) gives the verification layer real teeth. Your operational know-how becomes a durable, owned asset — portable across models, frameworks, and clouds — instead of something you rent from a single provider and lose when you switch.
The Howdex Codex is an open, schema'd catalogue of operational procedures with CRDT-based sync. One agent learns a procedure; every other agent can pull it — with its receipts attached, so a consuming agent knows whether to trust it. A fleet of agents stops rediscovering the same workflows in parallel and starts compounding a shared library of proven know-how. Specialist agents publish what they're best at; generalist agents reuse it. The commons is inspectable, governed, and owned by you — not a hosted black box.
This is where it gets unfair, in your favor.
Frontier models are brilliant and expensive — and worse, what they learn evaporates and stays theirs. Howdex changes the unit economics of running agents:
- Learn with the best. Run a frontier model (the teacher) on a hard task. When it succeeds, Howdex captures the trace and distills a deterministic, verified procedure — with receipts proving it worked.
- Publish to the registry. That procedure goes into the Codex: portable, inspectable, model-agnostic guidance. No weights, no prompts to leak — just the operational know-how.
- Execute with the cheapest. A smaller, cheaper, open model (the student) pulls the verified procedure from the registry and follows it — inheriting the teacher's hard-won competence on that task at a fraction of the inference cost, with no LLM in the memory loop at all.
Pay frontier prices once to learn. Reuse forever on models that cost a fraction. Because the procedure is deterministic guidance rather than model-specific weights, it travels across model families, frameworks, and clouds — and because it's receipt-backed, the student knows which procedures are safe to rely on. Cross-model portability is a core design goal of Howdex; large-scale independent transfer benchmarks across models are in progress.
The Codex enforces one rule: no proof, no procedure. A learned procedure is candidate until inspectable receipts show a task-relevant verifier passed, only then is it verified. Every entry carries provenance, failed-attempt memory (what to avoid), policy and staleness metadata, and source-exclusion controls.
howdex codex lint ./codex # structural + receipt integrity
howdex codex policy-check ./codex # policy & risk gates
howdex codex verify ./codex # re-check verifier evidenceThe Codex is not a prompt library and not an execution engine. In Howdex, procedures are guidance, not executable authority — a consuming agent verifies in its current environment before it acts. That boundary is exactly what makes Howdex safe to put in front of high-stakes work.
successful traces ─▶ learn ─▶ canonical procedure (+ provenance, receipts)
│
retrieve by task similarity
│
guidance ─▶ next agent succeeds faster, cheaper
Deterministic canonicalization turns differently-worded traces into one procedure. Retrieval is budgeted and inspectable. Guidance is operational facts — never pasted source — so reuse proves competence, not copying. Auditable abstraction: optional LLM proposals, deterministic trust. Optional, fully auditable LLM-assisted abstraction can propose semantic equivalences, but storage, verification, and trust status always stay deterministic and reversible.
Credibility is the product, so here is the honest boundary:
- Howdex does not, by itself, make autonomous execution production-safe; it supplies guidance and verification metadata, and the consuming agent remains responsible for verifying before it acts.
- The Docker benchmark above proves single-procedure transfer to fresh agents, not compounding across many independent traces or teams at scale.
- Cross-model portability is a design property; we do not claim independently proven live transfer across models.
- Our AWM-style comparison harness currently runs in dry-run mode; we do not claim that dry-run AWM-style harness results are live performance, and we make no claim of beating AWM, WebArena, or Mind2Web. We do not claim that Howdex has beaten the AWM paper.
candidateCodex entries and unaccepted abstraction proposals are not verified.
Howdex is increasingly built through its own procedural-memory loop: as we develop it, Howdex captures the build traces, learns procedures, and attaches receipts from passing test runs. These dogfood metrics are internal evidence only, they are not external users, adoption, traction, market validation. They show only that the loop runs on a real workload (our own). The reproducible benchmark above, not these internal metrics, is the external evidence.
pip install howdex-ai # or: pip install git+https://github.com/rossbuckley1990-hash/Howdex.gitDeep dives live in docs/: architecture, the Codex standard and protocol, MCP, framework adapters, governance, observability, and the benchmark methodology.
Apache-2.0. Own your loop.