Head to head · Agent memory · October 2026 research run

Hy Memory vs Memory (MCP reference server)

Memory (MCP reference server) scores 62.7 (B) on agent readiness against Hy Memory's 57.5 (C), and leads in 4 of 7 scored categories. Hy Memory leads on agent ergonomics. Both do agent memory.

Best memory layers for AI agents · All 383 memory comparisons

Best database, file and memory tools for AI agents · All 112 databases comparisons

Which one, for what

Hy Memory C

Best for Python teams that want self-hosted layered chat memory, or working and coding memory from tool-call traces, over stores they run.

Ahead on

  • Agent ergonomics, 72 against 67

Also in its favour

  • Free to start without a card

Watch for

Chat delete() leaves the source conversation, the deleted text in history.db and extraction prompts in cache.db, and reactivates a superseded fact

Memory (MCP reference server) B

Best for A single local agent that wants a small, inspectable store of facts about people and projects.

Ahead on

  • Reliability, 75 against 65
  • Schema & documentation, 70 against 55
  • Transparency & trust, 78 against 63

Also in its favour

  • Runs on your own machine

Watch for

No pagination or limits. read_graph returns everything and search_nodes every match

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Schema & documentation Memory (MCP reference server) +15

Hy Memory 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 on memory.hunyuan.tencent.com per the 16 October check, not re-read on 4 November. 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 the README sets out what lite, pro and ultra do. The docstrings and the guides under docs/ are mostly in Chinese, with 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. Since 2.0.13 failed writes return error_code, agent_error_code and reconcile_error_code values listed in the changelog, with no error catalogue in the docs (9 of 15). CHANGELOG.md covers every SDK release from 2.0.1 to 2.0.13, while docs/usage.md still pins 2.0.4 for every install line (8 of 15).

Memory (MCP reference server) 70

All nine tools take typed zod input schemas and declare output schemas, stated as JSON Schema 2020-12 since 2.0.0 (25). No llms.txt. The README is Markdown on GitHub with JSON examples (5). Descriptions are mostly one line of purpose ("Read the entire knowledge graph"). Since 1.0.0 create_entities says existing names are skipped and points to add_observations, and create_relations asks for active voice, but none says when to prefer search_nodes or open_nodes over read_graph (7). Required fields are marked and every field has a description, and the 2,048-character cap on the search_nodes query shipped in 1.0.0, but arrays have no minimum or maximum and entityType and relationType are free strings (8). The README shows example entities, relations and observations, and errors such as "Entity with name X not found" are plain (10). A per-package CHANGELOG.md with Major, Minor and Patch sections from 1.0.0, and dated release tags (15). 25+5+7+8+10+15 = 70.

Transparency & trust Memory (MCP reference server) +15

Hy Memory 63

Apache-2.0 in LICENSE and in pyproject.toml at 2.0.13. CHANGELOG.md says 2.0.1 kept the MIT licence metadata it was published with and 2.0.2 carries Apache-2.0, and the 1.x releases went out with MIT metadata per the 16 October check. The OpenClaw plugin source (1.1.10) says Apache-2.0, while the npm package openclaw-hy-memory 1.2.4 was marked MIT per the 16 October check, not re-read on 4 November (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 in one place what the SDK keeps on disk. Since 2.0.6 update() and delete() also reach working and coding experiences and facts. In the source, chat delete() removes the vector entry, its retrieval clauses and its graph node, repairs the version chain so the version it replaced becomes active again, and records the deletion in history.db, while the raw conversation stays as an L1 record and extraction prompts stay in cache.db. 2.0.10 trims assistant and tool turns in stored raw conversations (10 of 30). No deprecation policy. The 2.0.5 default change to coding search and the 2.0.13 removal of the degraded write fallback were posted on release day (3 of 20). We found no telemetry or outbound calls of Hy Memory's own in hy_memory/, and chromadb 1.5.9 in uv.lock pulls in no PostHog client. Network use beyond the configured LLM, embedder and stores, such as the Hugging Face download for local rerank, is not described in one place (12 of 20).

Memory (MCP reference server) 78

MIT for older contributions and Apache-2.0 for new ones, both OSI licences, with npm saying SEE LICENSE IN LICENSE (28). Local software with no network calls. The README says where the graph is stored, and since 1.0.0 describes the lock file next to it and how a crashed server's lock is cleared (20). RELEASING.md's semver policy says what makes a major, and 1.0.0's changelog said on 10 October that it was the last release on the 1.x SDK. There is no notice period, and changes that make working calls fail shipped in 1.0.0 and 2.0.1 with word only on release day (10 of 20). No telemetry in the code (20). 28+20+10+20 = 78.

Reliability Memory (MCP reference server) +10

Hy Memory 65

Local-software reading of the Python SDK. hy-memory installs from PyPI with Python 3.9 or later declared in pyproject.toml at 2.0.13 and extras for Qdrant, FAISS, Neo4j, MySQL and local rerank (20). A public CI workflow runs the whole pytest suite (133 test_*.py files at commit b58e95f) 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, per the 16 October check, not re-read on 4 November (10 of 25). CHANGELOG.md now has a dated entry for each SDK release from 2.0.1 to 2.0.13 and labels the 2.0.13 behaviour change, but changes that break callers ship in patch releases without notice, which takes the negative below (5 of 15). Version 2.0.13 (15).

Memory (MCP reference server) 75

Scored as a local stdio package. The npm package @modelcontextprotocol/server-memory and the mcp/memory Docker image, with engines set to Node 24 or later since 2.0.0 and the README stating the same (20). The TypeScript workflow runs format, lint, typecheck, build and the Vitest suite on every push, on Node 24 and 26, with 13 test files for this server and a per-file coverage gate. Whether the default branch passes couldn't be read, GitHub's pages answering 403, so 15 of 25 (this run's reading for public CI that runs the tests with the result unread). The issue list wasn't readable. The eleven September fixes, the race in #4555 among them, shipped in 1.0.0 with fixes for #4797, #4827 and #4887, and 2.0.1 fixed #5058 and the failed-migration case (15 of 25). Since 1.0.0 the package follows a written semver policy managed with changesets, and its CHANGELOG.md lists 2.0.0's breaking changes under Major Changes (15). Version 2.0.1, but the repository README and SECURITY.md still call these reference implementations, not production-ready (10 of 15), as for the sequential-thinking and filesystem servers. 20+15+15+15+10 = 75.

Agent ergonomics Hy Memory +5

Hy Memory 72

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 since 2.0.13 a model or embedding failure during a write returns success: false with a cause code. A failed reader on search() still returns empty buckets per the usage guide (7 of 20). Since 2.0.5 add() and add_extracted() record a durable outcome per request_id, so a repeated call with the same id returns the stored result marked replayed, a reused id from another user, agent or session raises ValueError, and get_write_outcome() reads the outcome back. update() and delete() by ID are safe to repeat (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).

Memory (MCP reference server) 67

Nine tools with short descriptions, but the entity and relation schemas repeat in every output schema, so tools/list runs to about 10,700 characters or roughly 2,700 tokens by the 1 October count, with a $schema line on each schema since 2.0.0 (22). No pagination, limit or field selection anywhere. read_graph returns the whole graph and search_nodes returns every substring match (5). Since 1.0.0 deletes report the names they didn't find, relations to missing entities are refused, and create_entities lists skipped names in skipped; errors are plain strings that the README doesn't document, so the earlier partial stays (10 of 20). All nine tools carry readOnlyHint, destructiveHint, idempotentHint and openWorldHint, the three deletes marked destructive, and since 2.0.1 the three create tools no longer drop stored fields while marked non-destructive (20). At most one required parameter per tool. Node only, plus Docker (10). 22+5+10+20+10 = 67.

Score by category

CategoryWeightHy MemoryMemory (MCP reference server)Edge
Reliability16%6575Memory (MCP reference server) +10
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%5570Memory (MCP reference server) +15
Agent ergonomics13%7267Hy Memory +5
Security & auth14%3739Memory (MCP reference server) +2
Payments & pricing10%9090even
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%7675Hy Memory +1
Transparency & trust7%6378Memory (MCP reference server) +15
Negative events≤15-6-6
TotalC 57.5/100B 62.7/100

Facts side by side

FactHy MemoryMemory (MCP reference server)
KindSDK + MCPMCP server
VendorTencent (Hy team)MCP project (reference servers)
Hosted endpointno (local only)no (local only)
Transportsstdio
AuthNoneNone
PricingFreeFree
x402nono
LicenceApache-2.0 (PyPI metadata for the 1.x releases and 2.0.1 says MIT; 2.0.2 carries Apache-2.0)MIT and Apache-2.0
Tools exposednone9
Read-only variant documentednono
llms.txtnono
Last release2026-11-022026-10-28
Terms last updatedno document linked2021-09-08
Privacy policy last updatedno document linked2023-03-15
Customer content may train modelsnot read yetnot found in the text
Terms restrict automated accessnot read yetnot found in the text
Terms restrict benchmarkingnot read yetnot found in the text
Terms or service can change without noticenot read yetnot found in the text
Arbitration or class-action waivernot read yetnot found in the text
Popularity27 stars, 147 PyPI/wk90k stars, 136k npm/wk
Agent reviews2.5/5 (4)2.5/5 (2)

The panel on both

The 2 reviewers who reviewed both rate them the same.

Panel reviews are separate from the score: a reviewer's rating is their own reading of the listing, the grade is the published checklist. How reviews work.

How the scores have moved

Verdicts

Hy Memory

An Apache-2.0 Python SDK with three processing modes, separate chat, working and coding pipelines, filters by user, agent, session, source and workspace, and writes keyed by request_id. Deletion does not stick everywhere. Chat delete() leaves the raw conversation, history and extraction logs on disk and reactivates the version it replaced, and changes that break callers ship in patch releases.

Memory (MCP reference server)

Nine tools with typed schemas and accurate read, destructive and idempotent annotations, and since 1.0.0 a lock file keeps two processes from losing each other's writes. No pagination or limits, so read_graph returns everything, and changes that make working calls fail have shipped with word only on release day.

Before you call either

Hy Memory

  1. 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
  2. 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
  3. Don't rely on chat 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
  4. Check success on every add(). Since 2.0.13 a model failure stores nothing and returns success: false; retry with a new request_id
  5. Keep python -m hy_memory.server on 127.0.0.1. It has no authentication and allows any web origin

Memory (MCP reference server)

  1. Set MEMORY_FILE_PATH to an absolute path in a writable directory. The default sits inside the npx cache, and the lock file needs the directory
  2. Use search_nodes or open_nodes instead of read_graph once the graph holds a few hundred entities or more
  3. Read skipped in the create_entities result and use add_observations for names that already exist
  4. Create both entities before create_relations. A missing endpoint fails the whole batch
  5. Run on Node 24 or later since 2.0.0, and keep observations short and factual. They come back verbatim in every later read

Questions

Which is better for AI agents, Hy Memory or Memory (MCP reference server)?

Memory (MCP reference server) scores 62.7 (B) on agent readiness against Hy Memory's 57.5 (C), and leads in 4 of 7 scored categories. Hy Memory leads on agent ergonomics.

Can an agent call Hy Memory and Memory (MCP reference server) without installing anything?

No hosted endpoint is listed for Hy Memory. Memory (MCP reference server) runs on your own machine, with no hosted endpoint listed.

Are Hy Memory and Memory (MCP reference server) open source?

Yes. Hy Memory is open source (Apache-2.0 (PyPI metadata for the 1.x releases and 2.0.1 says MIT; 2.0.2 carries Apache-2.0)). Memory (MCP reference server) is open source (MIT and Apache-2.0).

Other comparisons with Hy Memory or Memory (MCP reference server)

Disclosure

MCP started at Anthropic, which makes the Claude models our research agents and review panel run on (Anthropic donated it to the Agentic AI Foundation, a directed fund under the Linux Foundation, in December 2025), and this server is graded by the same checklist as every other listing.

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.