Hy Memory

# Fix list: Hy Memory

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/hy-memory, the October 2026 research run, assessed 16 October 2026. Grade C, 57.7 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Hy Memory: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Security & auth, 37 out of 100, up to 11 more on the total

Why it scored 37: Local-software reading, as for LangMem and Graphiti. The SDK holds no credential of its own and opens no port. LLM and embedder keys come from the environment or a `.env` file. Separation between users rests on the `user_id` the caller passes, and `search()` accepts several `user_ids` at once by design. The bundled `python -m hy_memory.server`, which the OpenClaw plugin starts, listens on 127.0.0.1 port 19527 with no authentication, sends `Access-Control-Allow-Origin: *` on every response and allows DELETE in its CORS preflight (14 of 30). No read-only mode. `delete_all()` removes all of a user's memories with no confirmation step, which only the OpenClaw plugin's `memory_delete` tool adds (`confirm: true` for bulk deletion) (4 of 20). Memories return to the model as stored, and we found no prompt-injection or memory-poisoning guidance (0 of 15). On by default, `history.db` records each add, update and delete with the memory's text, and the pipeline-trace table in `cache.db` keeps each extraction prompt and response, readable through `history()` and `MemoryInspector`. An operations audit table is off by default (12 of 15). SECURITY.md routes reports to GitHub private advisories, publishes no email address and adopts no response time. No security.txt on memory.hunyuan.tencent.com or www.tencent.com, and no bug bounty or certification named for the project. CI pins its actions by commit with read-only contents permission (7 of 20).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 2. Schema & documentation, 55 out of 100, up to 7.3 more on the total

Why it scored 55: Library reading. No OpenAPI or tool schema applies, and the SDK defines no agent tools. Public methods carry Python type hints for their arguments, but every result is an untyped `Dict[str, Any]` whose shape is described only in docstrings and guides (12 of 25). No llms.txt (memory.hunyuan.tencent.com/llms.txt returns 404). The guides are Markdown files in the repository that an agent can read raw (5 of 10). Docstrings on `add()` and `search()` say when each scene and search mode calls an LLM and when it doesn't, and the README sets out what lite, pro and ultra do. The docstrings and every guide under docs/ are in Chinese, with only the README in English (13 of 20). `mode`, `scene`, `search_mode` and `reader` are plain strings checked at run time, and `metadata` and `meta` are free-form dictionaries (8 of 15). Worked examples in the README and guides. Errors are partly documented, such as empty result buckets when the reader fails and `ValueError` for invalid arguments, with no error catalogue (9 of 15). CHANGELOG.md covers two SDK releases, and docs/usage.md still pins 2.0.1 for every optional extra (8 of 15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## 3. Reliability, 65 out of 100, up to 7 more on the total

Why it scored 65: Local-software reading of the Python SDK. `hy-memory` installs from PyPI with Python 3.9 or later declared and extras for Qdrant, FAISS, Neo4j, MySQL and local rerank (20). A public CI workflow runs the SDK's pytest suite (87 test files, integration tests excluded) on Python 3.12 only, on pushes to main and on pull requests. The Actions results weren't readable from our session, so a green default branch is not confirmed (15 of 25). GitHub's issue and pull request pages weren't readable either, so open crash reports were not counted. The public history is ten commits over two days, with fixes for timeout propagation and a Qdrant threading problem (10 of 25). CHANGELOG.md has dated entries for 2.0.1 and 2.0.2 only. The 34 PyPI releases of the 1.x line between May and July 2026, and the move to 2.0, have no notes on breaking changes (5 of 15). Version 2.0.2 (15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## 4. Agent ergonomics, 60 out of 100, up to 6.5 more on the total

Why it scored 60: Read as a library with the API lines. `search()` sizes its output with `limit` (default 10), `schema_limit`, `intention_limit` (default 0), `min_score` and `layers`, and returns content with a score in separate normal, schema and proactive channels, with no field selection (18 of 25). Filters on user, agent and session IDs, `include_sources` and `exclude_sources`, `created_after`, and `workspace_id` and `branch` for coding memory. `list_memories()` pages with `limit`, `offset` and `order`, while `search()` has no offset (18 of 20). Invalid arguments raise `ValueError` with a message, and `evolve()` returns `success: false` with an error. A failed reader returns empty buckets that look like no memories, and a failed routing or rewrite call falls back to chat and the original query, per the usage guide (7 of 20). `update()` and `delete()` by ID are safe to repeat. `add()` takes a `request_id`, which a code comment links to idempotent queue consumers, but we found no deduplication by it in the write path, so a retried `add()` runs extraction again and relies on reconciliation to merge duplicates (8 of 20). One constructor works with only `OPENAI_API_KEY` set, defaulting to gpt-4.1-nano, text-embedding-3-small and a local Chroma store, and every method has an async twin. Python only (9 of 15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 5. Payments & pricing, 60 out of 100, up to 5 more on the total

Why it scored 60: Apache-2.0 package the owner runs, with nothing to buy from Tencent, read with the self-hosted rule. 20 for pricing, 20 for a free start and 20 for use without an account. No x402, MPP or L402 in the docs or source (0). Embedding calls in every mode, and LLM calls in pro and ultra, are paid to whichever provider the owner configures.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 6. Transparency & trust, 62 out of 100, up to 3.3 more on the total

Made of editorial 50, provenance 73.

Why it scored 62: Apache-2.0 in LICENSE, in the 2.0.2 package metadata and for MemHub. PyPI metadata for every 1.x release and for 2.0.1 says MIT, and CHANGELOG.md says the repository adopted the Apache-2.0 text after 2.0.1 was published. The npm package `openclaw-hy-memory` 1.2.4 is still marked MIT while its source says Apache-2.0 (25 of 30). The README says only an LLM and an embedding endpoint are needed, and the usage guide says lite calls no extraction LLM while pro and ultra send content for extraction. Nothing says what the SDK keeps on disk. In the source, `delete()` removes the vector entry, its retrieval clauses and its graph node, but the raw conversation stays in the store as a hidden L1 record, `history.db` keeps the deleted text, extraction prompts stay in `cache.db`, and deleting a fact's newest version makes the version it replaced active again. No public method deletes working or coding memories (10 of 30). No deprecation policy. The environment reference says older settings no longer take effect, without dates (3 of 20). We found no telemetry or outbound calls of Hy Memory's own in `hy_memory/`. Network use beyond the configured LLM, embedder and stores, such as the Hugging Face download for local rerank, is not described in one place, and the SDK leaves Chroma's own settings at their defaults (12 of 20).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Status page: not found (0 of 10)
- security.txt: not found (0 of 10)

## 7. Maintenance & community, 76 out of 100, up to 2.1 more on the total

Why it scored 76: `hy-memory` 2.0.2 reached PyPI on 16 October 2026, the day of the check (30). Three releases in the last 90 days, 1.2.21 on 22 July, 2.0.1 on 15 October and 2.0.2 on 16 October, by PyPI's UTC upload times (20). The repository's first commit, 'Initial open source release', is dated 16 October 2026, and its history shows an outside contributor's timeout fix merged the same day as pull request 5. Issue and pull request pages weren't readable from our session, so reply times were not judged (10 of 25). One current official package, Python only. There is no MCP server, so the registry line does not apply (10 of 15). CI with pinned actions runs on every push to main. Core dependencies such as pydantic, numpy and chromadb are unpinned above a floor, the SDK has no lockfile, and 2.0.1 shipped with MIT licence metadata that 2.0.2 corrected (6 of 10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- Judgement call. Graded on the Python SDK (`hy_memory.HyMemoryClient`), as the brief asked. MemHub is a separate product in memhub/ with its own versions and is not graded.
- Judgement call. Not tagged `mcp`. The SDK and plugins/ ship no MCP server; plugins/README.md lists `mcp/` as reserved and the directory doesn't exist. The only MCP server in the repository is MemHub's recall gateway, which MemHub installs into each host.
- firstReleased is 2026-05-21, the upload of `hy-memory` 1.0.0 to PyPI, not the open-source release of 16 October. The website's 1.0 table says 1.0 was not open-sourced, while PyPI carried 1.x wheels with MIT metadata. The WeChat article the README links for 1.0 wasn't read.
- We did not install or run the SDK. Deletion behaviour comes from reading `hy_memory/client.py`, `hy_memory/pipelines/writer.py`, `hy_memory/data/` and `hy_memory/coding/store.py` at commit f79770c.
- unchecked: GitHub issues, pull requests, Actions results and whether private vulnerability reporting is switched on. WebFetch read the repository front page but refused other github.com pages as outside its allowed set, and curl to github.com is intercepted by the session's proxy.
- unchecked: the EMNLP 2026 acceptance of arXiv 2606.09483. Its arXiv page (v1, 8 June 2026) carries no acceptance note, and the EMNLP accepted-papers list wasn't read.
- unchecked: the ICLR 2027 submission of the CMB paper. The PDF in docs/papers names no venue, and OpenReview wasn't read.
- The technical report (arXiv 2610.22196, v1 on 15 October 2026) and the Hugging Face model tencent/Hy-Memory-30B-A3B (created 15 October 2026, tagged Apache-2.0, 51 safetensors shards) were read as metadata only. The model was not downloaded or run.
- Vendor claims, not verified. The website says Hy Memory ranks first among open-source frameworks on LoCoMo (72.6%), MemoryArena (55.31%), CodeMemoryBench and SWE-ContextBench, that Hy-Memory-30B-A3B averages 55.49 against 54.62 for 'Opus5', and that Hy Memory is the industry's first memory framework to propose working and coding memory. No deduction was taken for the 'first' claim, which the page doesn't define; a later pass could check it against earlier frameworks.
- unchecked: whether chromadb's own anonymised telemetry is on in the versions installed. The SDK passes no settings to `chromadb.PersistentClient`.
- WebFetch refused memory.hunyuan.tencent.com, pypi.org and the other hosts below as outside its allowed set, so the website and its 1.0 page, PyPI's JSON API, pypistats, the npm registry, the MCP registry, Hugging Face's model API, arXiv abstract pages, RDAP and the security.txt locations were read with curl. memory.hunyuan.tencent.com has no robots.txt (404), huggingface.co allows all, and arxiv.org allows /abs, whose 15-second crawl delay we kept.
- The website links github.com/Tencent-Hunyuan/MemHub for MemHub, while memhub/README.md names this monorepo as the source. The MemHub repository wasn't read.
- The npm package `openclaw-hy-memory` is at 1.2.4 (22 July 2026, MIT), while the source in plugins/openclaw is at 1.1.10 under Apache-2.0.

## Weaknesses

- `delete()` leaves the source conversation, the deleted text in `history.db` and extraction prompts in `cache.db`, and can reactivate a superseded fact
- No public method deletes working or coding memories. `delete()` and `delete_all()` touch only the chat store and the graph
- The guides under docs/ and the docstrings are in Chinese. Only the README is in English
- Every 1.x release and 2.0.1 went to PyPI with MIT licence metadata before the repository adopted Apache-2.0
- The bundled local HTTP server has no authentication and sends `Access-Control-Allow-Origin: *` on every response

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Pass `user_id` on every `add()` and `user_ids` on every `search()`. Users are kept apart only by the IDs you pass, and `search()` returns every listed user's memories
- Choose the mode before writing. `lite` makes no LLM calls and stores raw text. `pro` and `ultra` send what you store to the configured LLM and embedder for extraction
- Don't rely on `delete()` to erase a fact. The raw conversation, `history.db` and `cache.db` traces keep copies, and deleting a fact's newest version reactivates the older one
- Send tool-call messages with `scene="working"` or `scene="coding"`, and the same scene on `search()`. Without tool messages a write falls back to chat memory
- Keep `python -m hy_memory.server` on 127.0.0.1. It has no authentication and allows any web origin

## What the review panel asked for

- dated deprecation policy
- 1.x to 2.0 migration notes
- Typed result models
- English docstrings and guides
- distinct reader failure error
- English API reference
- authentication on the local server
- deletion that removes every copy

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.