Changelog and Release Status
Hy Memory and MemHub are versioned and released independently.
Unreleased
MemHub 1.0.6 — 2026-11-02
Build 37, on hy-memory==2.0.11. Linux x64/arm64 Runtime and macOS arm64
App and Runtime.
- Interfaces in 11 languages: English, Simplified and Traditional Chinese, Japanese, Spanish, German, Portuguese, French, Korean, Russian and Indonesian. Choosing any language other than Simplified Chinese also sets the SDK prompt language; Simplified Chinese keeps the automatic zh/en detection.
- Memory modes: lite (no LLM, keeps raw conversations), pro (LLM extraction)
and ultra (pro plus the knowledge graph and a nightly evolve run). Set with
GET/PUT /api/v1/settings/memory-modeormemhub settings mode. Raw conversations are now kept permanently.POST /api/v2/evolveandmemhub evolverun System 2 on demand. - Safe embedding-model switching: an embedding-space registry, a switch that
pauses writes, re-embeds every user's vectors and the graph, verifies and
commits atomically, and resumes after a crash or an App update. Every start
scans the vector store and merges data left in other spaces. When the
embedding service is unavailable, search returns
embedding_unavailable. - Multi-user Runtime: memory APIs and the CLI take
user_id(defaultlocal); records are isolated per user; Hermes profiles can be connected as separate users. The App keeps usinglocal. - Configurable SDK features through the CLI and API (
memhub settings features): graph rerank, episode recall, session episodes, episode summaries, intention and few-shot. Defaults are unchanged. - Search and recall pass through the SDK's search parameters with server-side
score floors; the memory library filters by layer;
/api/v2/rawreturns data again; v2 source and import routes; profile version history. - Upgrade fixes: migration no longer stops without a model profile (#234) or
on a pre-existing readback mismatch (#229, #240); source grants work after
the cutover (#233); reads use the active embedding (#242); the embedding
endpoint no longer follows the generation
base_url. - Hermes import drops gateway notices and probes and unwraps relayed
messages;
memhub logsno longer reports valid older records as damaged.
Hy Memory 2.0.11 — 2026-11-01
- Published
hy-memory==2.0.11. - LLM prompts now ship in 11 languages:
en,zh-Hans,zh-Hant,ja,es,de,pt,fr,ko,ruandid. They live in one registry,hy_memory/prompts/, with one module per language. Call sites useget_prompt(name, language), so a new language only needs a new module. The registry covers chat extraction (system prompt, basic_info, intentions, few-shot including Example 11), the no-intention variant, reconcile and its search-query prompt, the working/coding extractor, both coding reconcilers and the chain summary, the episode and topic summaries, System 2 evolve, and the user name line. The English and Simplified Chinese strings are unchanged byte for byte, and a test pins them. The old names (prompts_zh.*_ZH,extractor.EXTRACT_SYSTEM_PROMPT, ...) still import. The nine new languages are machine translations, and native-speaker review is recommended. - New optional
languageto choose the prompt language explicitly. Set it onMemoryConfig.language(envMEMORY_LANGUAGE) for a client default, or per call onadd/async_add,add_extracted,finalize_session,evolveandsweep. A per-call value overrides the config. The value goes through every LLM call of the request: extract and basic profile, reconcile, the coding path, the summaries and System 2. Aliases:zh,zh-CN→zh-Hans;zh-TW,zh-HK→zh-Hant;pt-BR→pt. An unknown code raisesValueErrorlisting the supported codes, and the HTTP server returns 400. With an explicit language, memories are written in that language even when the conversation uses another one. Without a language, the existing zh/en auto-detection runs unchanged and memories keep the source language.searchhas nolanguage: its judge and rewrite prompts are English-only and keep the query's language. - The HTTP server accepts
languageon/api/v1/addand/api/v1/add_extracted.
Hy Memory 2.0.10 — 2026-11-01
Published
hy-memory==2.0.10.New
hy_memory.reembedmodule for switching the embedding model. Writers must be paused for the whole run. Nothing is derived from the provider or endpoint: source and target are named explicitly, and the identity of a space is the model name plus the width (space_identity). Only Chroma and Kuzu are supported.vector_space_names(config, base_name=None)returns every physical collection a client opens for the six vector roles: memory, entities, coding keys and working facts, each in the working and coding variants.reembed_vector_space(source, target, embedding=...)pages every raw source collection, covering all users and all layers including L1_RAW. It embeds the text the write path embedded (reembed_text: the canonical state'sstate_retrieval_text, a schema'sretrieval_text, otherwise the document). It then writes each record into the target with the same id, document and metadata. Each batch's widths are checked before anything is written. A cancel stops at a batch boundary (ReembedCancelled). A rerun skips the ids the target already holds, so it resumes, and a merge never overwrites the target. The source is never written.plan_reembed,verify_reembedanddrop_vector_spacecomplete the vector side.verify_reembedchecks ids, metadata, per-user counts and the width, and runs a self-recall sample.drop_vector_spacerefuses collections that a live client in this process holds open.scan_vector_spaces(persist_directory)is read-only. It lists every space on disk with its roles, counts, stored width, per-user counts and the models recorded in metadata.diff_spaces(source, target)returns the ids that the source holds and the target lacks, whatever their status.reembed_graph(source_path, target_path, embedding=..., target_dims=...)builds a new Kuzu graph at the new width. It copies every node and edge and recomputesMemory.embedding,Memory.beh_embedding, and the Topic and HierarchyTopic embeddings, then rebuilds the HNSW index. The source graph is opened read-only and kept for rollback. A run that fails or is cancelled leaves no target behind, and a rerun starts over.verify_graph_reembedanddrop_graphcomplete the graph side.
EmbedService.embed_batchnow splits large batches into requests ofembedder.batch_sizetexts each (MEMORY_EMBEDDER_BATCH_SIZE, default 64); before, it sent them all in one request. On a 429 it waits for theRetry-Aftertime plus jitter, capped at 60 s. It takes a per-callmax_retries. A returned vector whose width differs from the configuredembedding_dimsnow raisesEmbeddingDimensionMismatchErrorand is not retried.Raw conversations are stored already trimmed. MemHub now keeps every raw forever and lite mode stores nothing else, so tool output and long pasted logs no longer stay in full. One rule (
hy_memory/utils/raw_trim.py) is applied to each message when the raw is written:userandsystemturns are kept in full;assistantturns and other speakers ([John]:) longer than 500 characters, after stripping, are cut to the first 500 plus.... This is the limit history injection used before;toolturns keep their head and tail, and the middle becomes[...N bytes omitted by truncate...], with the coding limitcoding.tool_result_max_bytes(MEMORY_CODING_TOOL_RESULT_MAX_BYTES, default 2048).
The rule is idempotent. It covers chat
L1_RAWin lite, pro and ultra (lite keeps tool turns, because it has no coding route) and thememory_historyADD record. The coding/workingcoding_tracewas already trimmed when it was written. Its tool and assistant truncations are now idempotent, so reading the history back no longer rewrites the omitted-count markers. The text embedded forL1_RAWis the trimmed text. New raws carrycustom.raw_trimmed=true.Extractor history ("last k messages") now joins the stored raws directly and does not trim them again. Raws written before this change have no
raw_trimmedflag, so they are still trimmed with the same rule when they are read. The extractor still receives the current batch in full (tool turns are still removed on the chat route).
MemHub 1.0.5 — 2026-10-28
Build 36, on hy-memory==2.0.8. Linux x64/arm64 Runtime and macOS arm64
App and Runtime. Upgrading from 1.0.4 migrates the existing library to SDK
authority on first launch.
MemHub turns off the HY pipeline log and trace at every diagnostic level (hy-memory 2.0.8
MEMORY_PIPELINE_LOG_ENABLED=false,MEMORY_PIPELINE_TRACE_ENABLED=false), so the per-step JSONL files underhy-memory/governed-v2/logs/and thepipeline_logsrows stop growing. The Debug panel no longer shows the HY pipeline section. Existing files are left in place.HY SDK text logs are written once, to
hy_memory.log. They are no longer copied toservice.logandservice-error.logthrough the SDK's stdout/stderr streams. SDK warnings and errors also go tocore.log, so the Debug panel still shows them. A live switch of the diagnostic level now also changes the level ofhy_memory.log.MemHub shows when working and coding experiences were talked about. After the SDK upgrade every one of them read "时间未知" (time unknown) because the SDK never stored their conversation time; hy-memory 2.0.7 stores it and Core now passes it through the memory list, detail, search and home summary. An experience stored before 2.0.7 shows its write time, labelled "写入于" / "写入时间", instead of "时间未知". Metadata repair also writes the conversation time it recovers from the migration bundle into the SDK's coding/working stores when they already have the column.
The SDK migration no longer fails its first run on a home with legacy vectors: reattaching them left a Chroma instance open with its own settings, and the readback that followed could not open the same store ("An instance of Chroma already exists ... with different settings").
MemHub "添加记忆" (add a memory) no longer waits behind unprocessed history. Notes are claimed ahead of the event-time window that keeps history strictly old to new; they still run one at a time and in order within their session.
The SDK memory page now reads its list, count, source and time labels and graph from one page response for one source filter, so rows, totals and the map no longer disagree. Memories with no recorded source are shown and filterable as "unidentified source"; library search keeps only memories that contain the words within the SDK's relevance window.
Metadata repair v2 follows
governed_memory_evidenceacross every pre-cutover bundle, which restores the source of memories the cutover backfilled without one, and moves impossible future conversation times back to the write time.Reused the existing three-column memory map for SDK-authority list/detail views and removed the retired "original evidence" surface. SDK-native status, source, conversation time and creation time now render without fabricated revisions; source and text filters are applied by the paged Core endpoint.
SDK memory search now calls the SDK search contract instead of listing the entire library in Core. Corrupt empty records are skipped and diagnosed without mutating data or failing a page, and post-cutover metadata repair restores source and original times from the verified migration bundle.
Kept Cursor's WAL change token separate from the file mtime so a hash can no longer appear as a conversation date centuries in the future.
A trusted packaged Runtime now repairs a dirty same-build installation by staging a clean replacement, and
runtime_service_not_readyis rendered as user-facing startup guidance instead of an internal code.Fixed legacy-to-SDK startup migration to adopt native Coding Experience and Working Fact IDs instead of importing them as Chat layers. Raw migration now rebuilds bounded user/assistant turns from the latest three days, excludes tools, and enforces the same rolling retention after cutover.
Paused host capture writes with a live Desktop migration fence; Cursor History catch-up resumes from its existing checkpoint after cutover.
Removed MemHub's private
hy_memory_bridgeimplementation. MemHub now calls publicHyMemoryClientmethods at runtime; prompt/reconcile/response monkey-patches and private VDB/Coding runtime access were deleted. Kuzu Graph projection/admin and stopped-writer App migration are the documented product exceptions.Added a MemHub-managed native Hermes MemoryProvider with automatic prefetch, persistent per-session memory-ID deduplication, authorized Broker reads, and recall-context stripping to prevent memory recursion.
Added an extensible MemHub interface-localization layer with Simplified Chinese, English, and Spanish catalogs shared by Web and Electron-native surfaces. Interface language remains independent from memory content and model processing language.
Fixed governed-memory revisions and repeated evidence links failing when the visibility dirty-marker trigger encounters the same memory twice. Startup now replaces the affected legacy trigger without changing schema versions or governed and visibility data.
Kept post-write authority verification failures available for read-only reconciliation while preventing automatic or generic confirmed retries from invoking the model again. Automatic reconciliation skips those sticky rows before applying its batch limit, and linked durable jobs exclusively own retry scheduling for their memory tasks.
Removed MemHub's task-scoped HyPermission store wrappers and global model-content consent switch. The trusted embedded Hy Memory SDK now uses its native local stores directly.
Existing installations ignore the retired config flag, readmit previously parked pre-invocation work, and drop only the obsolete
hy_node_permissionsproof table. Evidence, governed memories, model profiles, SDK stores, and source/host grants are preserved.
Hy Memory 2.0.9 — 2026-11-01
- Published
hy-memory==2.0.9. - Placeholder values no longer overwrite the user's basic profile (L0). A
small extraction model filled
basic_infowith{"name": "用户"},The userorN/Awhen the user had said nothing about themselves, and each placeholder superseded the real value, so the L0 chain read 费天翔 → Fei → 用户 → The user. Two defences:upsert_basic_profiledrops placeholder values: empty,用户,the user,user,n/a,na,none,null,unknown,未提供,未知,不详,-,?and the field's own name, compared case-insensitively with whitespace and punctuation stripped. When nothing is left it writes no new version and no new L0 head. Placeholders already stored in the chain no longer override earlier real values when the profile is assembled, soget_user_profileof an affected user shows the last real value again without any migration. This also stops an assistant quoting an old memory ("The user's name is 用户.") from being written back.- The extraction prompts (zh and en) state the rule: only what the user
said in a
userturn, never the field name or a placeholder,{}otherwise, with a small-talk counter-example.
Hy Memory 2.0.8 — 2026-10-28
- Published
hy-memory==2.0.8. - New
MEMORY_PIPELINE_LOG_ENABLED(defaulttrue) turns the pipeline JSONL log ({MEMORY_LOG_DIR}/<subdir>/<date>.log) off. Withfalseno JSONL file or directory is written. WhenMEMORY_PIPELINE_TRACE_ENABLEDis alsofalse,store_pipeline_logbecomes a no-op that returnsTrueat once: no timing, redaction, background task or database write. The init log line now states whether the JSONL log is on or off. - Conversation and memory text no longer appears in INFO or WARNING logs. The
[add]input, search queries and rewrites, coding experience tasks, fact statements and subjects, reconcile op details, extracted info, embedding inputs and unparsable model output moved to DEBUG. INFO lines keep ids, counts and character lengths; parse-failure warnings remain but report only the length of the output, e.g.failed to parse output (812 chars). The[embed]INFO line now carriesrequest.input_charsinstead ofrequest.input. TheJSON_PARSE_FAILEDerror message from extraction no longer embeds the model output either.
Hy Memory 2.0.7 — 2026-10-28
- Published
hy-memory==2.0.7. - Coding and working experiences and facts now keep their conversation time.
add(..., memory_at=...)reached only the chat path before; the productivity path dropped it, so every experience had only its write time and MemHub showed them as "时间未知" (time unknown).coding_memory_metaandworking_factgain a nullablememory_atcolumn (epoch seconds) in both the working and coding stores; existing databases get it on open, and their old rows stayNone. Withoutmemory_ata write uses the time the request arrived, as chat does. When an experience is updated by a later conversation it keeps the later of the two times. list_memories(scene="coding","working"and"all"),getandsearchreturnmemory_atfor experiences and facts in the same field and format as chat (epoch seconds,Nonewhen unknown). Experience search results now also carrycreated_at.
Hy Memory 2.0.6 — 2026-10-27
- Published
hy-memory==2.0.6. search(..., rerank=False)skips every reranking stage for that one call (the local cross-encoder, the memory-graph rerank, and the BM25 rerank of working/coding facts), so results come back in retrieval-score order. The flag travels on the read request, so concurrent searches do not affect each other. The default (True) keeps current behavior;/api/v1/searchaccepts it too.- MMR diversity selection now runs only when the requested
limitis below 20 (MMR_MAX_TOP_K) and is skipped for larger limits, where its quadratic token-overlap pass is slow and gains little. Previously it ran for every limit.MEMORY_HYBRID_V2_MMR_LAMBDAstill sets the weight. Near-duplicate demotion follows the same gate: it compares each candidate with every earlier non-duplicate, which is quadratic in a pool that grows withlimit, and large result pages should show similar items anyway. search(scene="all")searches chat, working and coding at the same time and merges them with Reciprocal Rank Fusion (k=60) over each scene's own ranking, so no scene is crowded out by the other scenes' score scales. Each result says which scene and kind it came from. A scene that fails or is turned off is reported inerrorsand does not fail the search. Source filters apply to all three.list_memories(scene="all")lists chat, working and coding memories together, newest update first, withlimit/offsetpaging over the merged order andtotalsumming all three.scene="coding"and"working"now honorinclude_sources/exclude_sources, andtotalcounts only the filtered memories.update()anddelete()now work on working/coding experiences and facts, not just chat memories. They take the same optionalkind/variantasget(), and an ID found in more than one place raisesAmbiguousMemoryIdErrorinstead of guessing. Updating an experience replaces its solution and keeps its task, unless a newtask=is also passed; updating a fact replaces its statement. Deleting removes the stored row and its vectors.list_memories(scene="chat" | "coding" | "working")lists one kind of memory.codingandworkinglist that scene's whole store, newest update first (workspace_idstill narrows it), andtotalis now the real count rather than the page size. The newmemory_stats()returns the count and the latest created/updated time for each scene and their total, without reading any memory text. Previously coding and working memories could not be listed or counted without a workspace, so a library of 45 memories listed 22.- Reconcile keeps time order. Both reconcile prompts now state that only
memory_atdecides which memory is newer, that an older memory must never supersede or overwrite a newer one, and that an unknown time is not newer; the Chinese prompt now listsmemory_atamong the new-memory fields. The reconciler also enforces this in code: a SUPERSEDE or UPDATE aimed at a memory with a latermemory_atbecomes an ADD, so the newer memory stays current. - The HTTP server returns 409
ambiguous_memory_idand 400 for a badkind/variantinstead of 500, passeskind/variant(andtaskon PUT) to get/update/delete, and no longer reads a PUT/DELETE query string as part of the memory id. memory_stats()reads working/coding times, which are stored as naive UTC, as UTC rather than local time.
Hy Memory 2.0.5 — 2026-10-22
- Published
hy-memory==2.0.5. - Extended
get()across Chat, Coding Experience, Working Fact, and raw IDs; the returnedlayer,kind, andvariantidentify the actual record. - Added explicit raw retrieval through the existing
search()andlist_memories()contracts, including user, agent, session, and source filters. - Added public embedding-service injection and corrected extraction language detection for structured OpenAI-style messages.
- Added lifecycle status filters, durable idempotent write outcomes, and quiesced local data-plane snapshot/restore.
- Kept legacy migration mechanics out of the public Client API; separate migration-only raw/import methods from development previews were removed.
- Physical delete now reports partial VDB/Graph cleanup as failure.
- Coding search no longer injects the version-evolution diff block by default;
opt in with
MEMORY_CODING_CHAIN_BLOCK_ENABLED=true. - Serialized SDK SQLite cache/history connection lifecycle and tightened Coding reconciliation to its agent/session/workspace authority.
- Clarified persistent caller context through
meta, whilemetadataremains legacy processing configuration.
Hy Memory 2.0.5.dev4 — 2026-10-21
- Development preview used for SDK-authority migration validation; not the
final
2.0.5release. - Published as
hy-memory==2.0.5.dev4.
Hy Memory 2.0.5.dev3 — 2026-10-21
- Development preview; not the final
2.0.5release. - Published as
hy-memory==2.0.5.dev3. - Lets the existing
searchAPI explicitly query SDK raw withlayers=["l1_raw"], including existing source filters.
Hy Memory 2.0.5.dev2 — 2026-10-21
- Development preview; not the final
2.0.5release. - Published as
hy-memory==2.0.5.dev2. - Adds exact, idempotent semantic-memory import for legacy MemHub cutover without extraction, reconciliation or identity changes.
Hy Memory 2.0.5.dev1 — 2026-10-21
- Development preview for the MemHub SDK-authority migration; this is not the
final
2.0.5release. - Published as
hy-memory==2.0.5.dev1. - Adds lifecycle status filtering, durable idempotent write outcomes, exact raw APIs, offline local data-plane snapshot/restore, and fail-closed physical delete receipts.
Hy Memory 2.0.4 — 2026-10-17
- Published
hy-memory==2.0.4as a maintenance release.
Hy Memory 2.0.3 — 2026-10-17
- Added an optional runtime HTTP transport factory so embedded products can enforce request budgets without wrapping or restricting SDK local-store access.
- Published
hy-memory==2.0.3with unchanged memory extraction, reconciliation, storage, and Coding/Working behavior.
Hy Memory 2.0.2 — 2026-10-17
- Published
hy-memory==2.0.2with the current SDK fixes and Apache-2.0 package metadata.
Hy Memory 2.0.1 — 2026-10-16
- Published the Python distribution as
hy-memory==2.0.1. - The repository subsequently adopted the Hy-Memory Apache-2.0 license text.
Because PyPI release metadata is immutable, the next SDK patch release must
carry the updated license metadata;
2.0.1retains its published metadata.
MemHub 1.0.5 — Unreleased
- Updated the source version to
1.0.5(build 14). - Pinned the public SDK to the stable
hy-memory==2.0.5release.
MemHub 1.0.4 — 2026-10-18
- Updated the source version to
1.0.4(build 13). - Published the
v1.0.4source and release metadata.
MemHub 1.0.3 — 2026-10-17
- Updated the source version to
1.0.3(build 7). - Pinned the public SDK to
hy-memory==2.0.3. - Removed the HyPermission store wrappers and global model-content consent switch; the trusted embedded SDK now uses native local stores.
- Prevented uncertain writes from blocking later source work and added one reconciled automatic retry after ten seconds.
- Added product-facing memory pipeline counters and the log troubleshooting runbook.
- Allowed a verified packaged Desktop Runtime from a different commit to take over an equal-build install atomically while preserving downgrade and tamper refusal.
MemHub 1.0.2 — 2026-10-17
- Updated the source version to
1.0.2(build 6). - Pinned the public SDK to
hy-memory==2.0.2. - Pinned
onnxruntime==1.23.2to retain macOS x64 packaging support. - Fixed relaunch and upgrade failures when a prior managed Runtime exists.
- Embedded canonical build identity when an external packaging pipeline omits release environment fields.
- Added automatic Linux Runtime upload to tagged Draft Releases.
MemHub 1.0.1 — Superseded before release
- The candidate tag was not published as a GitHub Release;
1.0.2replaces it after the managed-Runtime relaunch fix.
MemHub 1.0.0
- Published the initial GitHub Release with macOS arm64 and Linux Runtime artifacts.