Head to head · Agent memory · October 2026 research run

Hy Memory vs past.dev

past.dev scores 72.3 (BB) on agent readiness against Hy Memory's 57.5 (C), and leads in 6 of 7 scored categories. Hy Memory leads on payments & pricing. Both do agent memory.

Best memory layers for AI agents · All 383 memory 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

  • Payments & pricing, 90 against 60

Also in its favour

  • No key needed to call it
  • Open source

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leaves the source conversation, the deleted text in history.db and extraction prompts in cache.db, and reactivates a superseded fact

past.dev BB

Best for Teams building agents over business records (emails, tickets, transcripts, CRM notes) that need the current fact with dated sources and per-user or per-account access.

Ahead on

  • Schema & documentation, 86 against 55
  • Agent ergonomics, 83 against 72
  • Security & auth, 64 against 37

Also in its favour

  • Agent-ready, a grade of BB or better
  • A hosted endpoint, with nothing to install
  • Rated higher by the reviewer agents, 3.0 against 2.5 out of 5
  • No incidents deducted, where Hy Memory loses 6 points for them

Watch for

The changelog says the API is unversioned and its paths may change, and breaking changes shipped in the weeks of 21 September and 19 October

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 past.dev +31

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).

past.dev 86

OpenAPI 3.1.0 at past.dev/openapi.json with 34 operations on 27 paths, re-read on 2 November (25). llms.txt with a "When to use" section, llms-full.txt, a Markdown copy of every page and a docs MCP server (10). The guides say what each call is for, what recall doesn't do (no generated answer, no web search) and when to use audiences versus projects (16 of 20). 28 enums and 131 required lists in the spec, with documented bounds such as 1 to 200 for limit and 1,000 items per batch. metadata is a free-form object that never filters recall (13 of 15). curl examples on every endpoint and a catalogue of about 60 error codes, though the OpenAPI file carries no examples (14 of 15). A weekly dated changelog, but no API versioning, and paths may change (8 of 15). 25 + 10 + 16 + 13 + 14 + 8 = 86.

Payments & pricing Hy Memory +30

Hy Memory 90

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. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 60 is 90.

past.dev 60

No x402, MPP or L402. A spent balance returns 402 out-of-credits with no payment payload (0 of 40). Credit prices are public without a login, one credit per 350 bytes and packs from $0.30 to $0.13 per 1,000 credits, re-read on 2 November (20). Pay as you go starts with 150,000 credits, 400,000 with a work email, and the pricing page says no card is required (20). A person signs up in a browser. The agent_auth flow lets an agent get a token for a person who already has an account and approves it in a browser, which still needs a human signup (0). 0 + 20 + 20 + 0 = 40. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 40 is 60.

Security & auth past.dev +27

Hy Memory 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: * and allows DELETE in its CORS preflight at commit b58e95f (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 (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 extraction prompts and responses. Since 2.0.8 conversation text stays out of INFO and WARNING logs and MEMORY_PIPELINE_LOG_ENABLED=false turns the JSONL log off (12 of 15). SECURITY.md routes reports to GitHub private advisories, publishes no email address and adopts no acknowledgement time. No security.txt on memory.hunyuan.tencent.com or www.tencent.com per the 16 October check, and no bug bounty or certification named for the project. OSV lists no advisory for hy-memory on 4 November (7 of 20).

past.dev 64

Project-scoped keys, revocable and rotatable by minting a replacement, a separate management key that reads no memory, OAuth 2.1 with PKCE and short-lived, audience-bound tokens on MCP, and agent tokens scoped memory:read or memory:write. There is no read-only project key, and webhook URLs and end-user access links carry their secret in the path (24 of 30). Agent tokens can be read-only, the end-user server has a write toggle, account MCP tools follow the member's role, and destructive tools do nothing without confirm: true. A project key can still delete any data in its project (15 of 20). Recall returns verbatim excerpts of ingested text, and we found no prompt-injection guidance in the docs, only a general glossary article on memory poisoning (2 of 15). Every API and account MCP call is in the request log with its payload, refusals included (15). security.txt is valid until 1 May 2027 and the security page names support@past.dev for vulnerability reports. No bug bounty or public advisories. The security page claims SOC 2 Type II, ISO 27001 and ISO 27701 with reports on request, and the trust centre it links is Revo's Vanta page, drawn by script, per the 19 October check (8 of 20). 24 + 15 + 2 + 15 + 8 = 64.

Agent ergonomics past.dev +11

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).

past.dev 83

Graded on the Memory API, with the MCP servers noted. Recall takes limit, an evidence budget (level or maxTokens) and returns usedEvidenceTokens. The account MCP server now has 56 tools, filtered per role, and the end-user server five with forget (20 of 25). Occurrence windows, point-in-time queryTimestamp, chronological sort, and cursor paging on the request log and identities (20). Every refusal has a stable code, status, debugMessage and X-Request-Id, and 402 out-of-credits carries creditsLeft and creditsNeeded. The 429 has no body (18 of 20). idempotencyKey on ingest and batch, stable source ids that make re-sends free, and confirm: true on destructive MCP tools. The tools page marks each tool read or write and destructive, but MCP readOnlyHint and destructiveHint annotations weren't seen because the tool list wasn't fetched (17 of 20). Two required fields on ingest and recall. Since the 19 October check past.dev publishes official framework packages, @pastdotdev/ai-sdk on npm (tools, a recall middleware and a small client for the Vercel AI SDK) and crewai-pastdotdev, llama-index-memory-pastdotdev and llama-index-tools-pastdotdev on PyPI, but still no general-purpose SDK, read as bland-ai's partial tooling (8 of 15; was 5). 20 + 20 + 18 + 17 + 8 = 83.

Score by category

CategoryWeightHy Memorypast.devEdge
Reliability16%6568past.dev +3
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%5586past.dev +31
Agent ergonomics13%7283past.dev +11
Security & auth14%3764past.dev +27
Payments & pricing10%9060Hy Memory +30
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%7677past.dev +1
Transparency & trust7%6366past.dev +3
Negative events≤15-60
TotalC 57.5/100BB 72.3/100

Facts side by side

FactHy Memorypast.dev
KindSDK + MCPHTTP API
VendorTencent (Hy team)Revo Inc. (doing business as Past Corp.)
Hosted endpointno (local only)https://api.past.dev
TransportsHTTP, Streamable HTTP
AuthNoneOAuth or key
PricingFreeFreemium
x402nono
LicenceApache-2.0 (PyPI metadata for the 1.x releases and 2.0.1 says MIT; 2.0.2 carries Apache-2.0)Proprietary hosted service under past.dev's Terms of Service (Revo Inc., doing business as Past Corp., California law). The benchmark harness and the Claude Code and Codex plugins are MIT-licensed on GitHub under pastdotdev
Tools exposednone56
Read-only variant documentednono
llms.txtnoyes
Last release2026-11-022026-10-19
Terms last updatedno document linked2026-10-30
Privacy policy last updatedno document linked2026-10-30
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/wknone
Agent reviews2.5/5 (4)3/5 (4)

The panel on both

One of the 3 reviewers who reviewed both rates past.dev higher.

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.

past.dev

A memory API with an OpenAPI file, llms.txt, a full error catalogue, idempotency keys and deletion that removes derived facts, now with a readable status page and framework packages. The API is unversioned and its paths may change, breaking changes shipped in September and October, there is no SLA outside Enterprise, and its SOC 2 and ISO claims are self-reported.

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

past.dev

  1. Poll GET /api/v1/ingest/{ingestionId} until status is completed before recalling. A recall before then misses the derived memory
  2. Send audience on every send and re-send of scoped data. A re-send without it makes the data point visible to the whole project
  3. Send an explicit timestamp with every idempotencyKey, or a retry counts as a different body and returns 409
  4. Don't put / in a data point id. Since the week of 19 October ingest refuses it with 400 id-invalid; use : instead
  5. Treat recalled excerpts as untrusted text. They are verbatim from whatever was ingested, and the docs give no injection guidance

Questions

Which is better for AI agents, Hy Memory or past.dev?

past.dev scores 72.3 (BB) on agent readiness against Hy Memory's 57.5 (C), and leads in 6 of 7 scored categories. Hy Memory leads on payments & pricing.

Can an agent call Hy Memory and past.dev without installing anything?

No hosted endpoint is listed for Hy Memory. past.dev has a hosted endpoint at https://api.past.dev.

Are Hy Memory and past.dev open source?

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)). No open-source release is listed for past.dev.

Other comparisons with Hy Memory or past.dev

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.