Experience reconcile no longer falls back to ADD-all, and fact reconcile no
longer inserts without superseding, when the model call fails, returns empty
content, or returns a reply that is not a JSON object with an ops list. The
write returns success=false, agent_error_code=RECONCILE_FAILED and the
cause in reconcile_error_code (LLM_UNAVAILABLE / EMPTY_RESPONSE /
JSON_PARSE_FAILED), the same as chat.
Each reconciler is split into plan() (retrieve + model decision, no store)
and apply() (store). CodingWriter plans both layers before applying either,
so a decision failure in either layer stores no experience or fact and
supersedes no older fact. reconcile() still runs plan + apply.
An experience reply of {"ops": []} is a valid "no changes" decision and
stores nothing. A non-empty ops list with no recognizable op is a parse
failure. A fact reply of {"ops": []} inserts without superseding.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Reconcile (add / add_extracted): when the reconcile model call fails,
returns an empty 200 reply or an unparseable reply, or reconcile raises
before anything is stored, the write no longer falls back to adding every
extracted memory as-is with success=true. The ADD-all fallback and its
degraded flags are removed. The write returns success=false,
agent_error_code=RECONCILE_FAILED and the cause in reconcile_error_code
(LLM_UNAVAILABLE / EMPTY_RESPONSE / JSON_PARSE_FAILED /
RECONCILE_EXCEPTION). Nothing new is stored, the episode counter does not
advance, and add rolls back its L1 raw (retry_safe=true), so a re-run with
a new request_id is clean. Only a literal [] still means "no changes".
RECONCILE_ENABLED=false keeps writing directly.
Working / coding path: an extraction model failure (LLM_ERROR,
EMPTY_RESPONSE, JSON_PARSE_FAILED) and an experience or fact embedding
failure (EMBEDDING_FAILED) now return success=false instead of success
with 0 ops or a node without a vector. Embeddings are computed before the
SQLite row is written and before an older fact is superseded. A failed
write does not append its trace or record chain rounds. A model that
answers with nothing to remember is still a success.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
When intention_dedup_enabled is off, no intention dedup should run, but
the writer's batch dedup still dropped near-duplicate intentions from
one extraction. Gate it on the same switch. Its threshold stays on
MEMORY_DEDUP_THRESHOLD.
Also skip the embedding backfill when dedup is off, since only dedup
uses it, and fall back to the env default on an invalid configured
threshold so a bad value cannot fail the read.
Refs #241
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Add MemoryConfig.recall.intention_dedup_enabled (default true, env
MEMORY_INTENTION_DEDUP_ENABLED) and recall.intention_dedup_threshold
(env MEMORY_INTENTION_DEDUP_THRESHOLD, falling back to the shared
MEMORY_DEDUP_THRESHOLD, then 0.95). The reader passes both from the
client config to recall_intentions. When off, read-time intention dedup
is skipped entirely: nothing is hidden and nothing is deleted.
Refs #241
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Move every LLM prompt that had a zh variant into one registry,
hy_memory/prompts/, with one module per language (en, zh-Hans, zh-Hant,
ja, es, de, pt, fr, ko, ru, id). Call sites look prompts up with
get_prompt(name, language), so a new language needs only a new module.
The en and zh-Hans strings are moved verbatim, a test pins their sha256,
and the old names (prompts_zh.*_ZH, extractor.EXTRACT_SYSTEM_PROMPT, ...)
still import.
Add an optional language: MemoryConfig.language (env MEMORY_LANGUAGE)
as the client default, plus a per-call language= on add / async_add,
add_extracted, finalize_session, evolve and sweep. It goes through
extract and basic profile, reconcile and search query, the coding
extractor and both coding reconcilers with the chain summary, the
episode and topic summaries, and System 2. Codes are validated, with
aliases (zh, zh-CN, zh-TW, zh-HK, pt-BR); an unknown code raises
ValueError. With an explicit language, memories are written in that
language. Without one, the zh/en auto-detection is unchanged. The
server accepts language on /add and /add_extracted.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Add hy_memory.reembed for switching the embedding model while writes are
paused. Source and target are named explicitly; a space's identity is the
model name plus the width.
- vector_space_names, plan_reembed, reembed_vector_space (resumable,
cancellable at batch boundaries, widths validated before any upsert, target
wins on existing ids, source never written), verify_reembed,
drop_vector_space.
- scan_vector_spaces and diff_spaces for MemHub's startup reconcile of stray
spaces.
- reembed_graph builds a new Kuzu DB at the target width, copies every node
and edge, recomputes Memory/beh/Topic/HierarchyTopic embeddings and rebuilds
the HNSW index; the source graph is opened read-only. verify_graph_reembed
and drop_graph.
EmbedService.embed_batch now chunks to embedder.batch_size (default 64),
honours Retry-After on 429, takes a per-call max_retries, and rejects vectors
whose width differs from embedding_dims.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
MEMORY_PIPELINE_LOG_ENABLED (default true) turns off the pipeline JSONL
log. With the DB trace also off, store_pipeline_log becomes a no-op that
does no timing, redaction, background work or database write. The init line
states the real on/off state.
Lines that carried conversation or memory text at INFO/WARNING now log ids,
counts and lengths; the text goes to DEBUG: [add] input, search queries and
rewrites, coding tasks and fact statements/subjects, reconcile op detail,
extracted info, embedding inputs and unparsable model output.
Bump hy-memory to 2.0.8 (not yet published).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The install blocks still asked for hy-memory 2.0.3 (and 2.0.1 extras),
MemHub 1.0.3 (build 7), and said memhub-cli was unpublished. pyproject.toml
ships 2.0.4, the MemHub workspace ships 1.0.4 (build 13), and README already
announces memhub-cli 1.0.5.
Update the install commands and the release metadata in the root READMEs,
the usage/search guides, the monorepo policy, and the MemHub docs. Historical
CHANGELOG entries are deliberately unchanged.
Co-authored-by: qiye.sjh <shijiahao_0802@youzan.com>
alvinfeiandClaude Opus 5.5
u7-u7andqiye.sjh