Repository navigation
Releases
Releases · itpartypattaya/hermes-plugin-holographic
Holographic 0.28.0
Added
- Wiki pointer facts carry entities,
#93 by @magicbluesmoke — thank you.
Bulk wiki ingest writes a page as a fact whose pointer line
(Wiki page: ai-agents/entities/llama-cpp.md) has no capitalised phrase, so the extractor found nothing and a
re-ingested page was never a topic of its own: it was reachable only by the words of its body. Such a line now
links its namespace (ai agents) and page slug (llama cpp); a section directory and structural slugs (index,
schema,log,readme,overview,changelog,notes,todo) are dropped, and a path with no namespace is
ignored. A namedelete_entityremoved is not recreated from a path — text never lifts a tombstone. Existing
stores get the links on the first writer open:_ENTITY_RULESis 5, so the index is rebuilt under the new rules
(up to 2,000 facts and 500,000 characters; a larger store onfact_store rebuild_entities).
Only the pointer shape is read — a line that starts with the[WIKI HINT …]marker beforeWiki page:— so an
ordinary note sayingWiki page: https://en.wikipedia.org/wiki/Foonames nothing, a protocol-relative URL
(//example.com/…) names nothing either, a leading byte-order mark does not hide the line, and Markdown code is
blanked out first, so a fact that shows the format in backticks or a fenced block stays documentation — while a
pointer on the line right after a closing fence is still read. Any writer may
produce a pointer line: it is a plain fact body (seedocs/wiki-pointer-entities.md).
Fixed
- Recall before a turn reads what the person wrote, not what Hermes wraps around it,
#96. The core handsprefetchthe turn as
the gateway built it: a Telegram reply carries the whole quoted message first, an attachment a note with its saved
path and instructions to the model, a group message the author's[Name], a cron run the fixed[IMPORTANT: You are running as a scheduled cron job…]hint and the run time, and a finished delegation or background process
arrives as a turn of its own. Measured on a live deployment (18 days, 103 recalls): replies were 5 % of messages but
43 of 70 Telegram recalls, and in 32 of them a fact came only from the quote; the author's name brought the same
fact in 14 of 20 group turns; the cron hint pushed facts the job needed out. Now the leading quote, attachment notes
(document, audio, video; an image's description stays, it is what was sent), the author's own name prefix and the
cron hint are left out of the search, a notice turn recalls nothing, and the 4000-character window starts at the
person's words. The presence gate still ties the turn by the whole message. A bare "ok" under a quote now recalls
nothing; the quote is in the model's view anyway.tests/test_holographic_service_prompts.py.
Holographic 0.27.0
Added
- French auto-extraction (
auto_extract, off by default), after
#86 by @maxandrieux — thank you. A sentence
that opens with a preference (Je préfère…,J'aime…,J'adore…,J'utilise…,Je n'aime pas / plus…,Je n'utilise pas…) is stored asuser_pref, one that opens with a decision made together (Nous avons décidé…,On a choisi…) asproject, at the same trust and through the same screens as English. Replies to the assistant are
skipped whatever adverbs they carry (Je n'aime pas du tout ça, merci,J'aime aussi cette proposition,Je préfère celui-ci pour le moment,On a choisi ça), while an object after a demonstrative is kept (J'utilise ce logiciel depuis dix ans,Je préfère celui de Paul).J'ai décidé…is a one-off, as "I decided" is in English;
questions,si…,que…(an instruction) and lead-ins (Moi, je préfère…) are left out. A reply followed by a
French preference in the same sentence keeps the sentence (J'aime ça, je préfère le thé vert), as English does. - Memory graph (retrieval Phase 4,
graph_expansion, off by default),
#91 by @magicbluesmoke — thank you. Two
entities a fact links get aco_occursedge inentity_relations(thetrg_fact_entities_edgetrigger, created by
a writer's_migrate()), and search walks ONE hop out of the
entities the query resolves to. A fact about an entity next to the query's is then reachable without sharing a word
with it (docs/memory-graph.md).
Bounded, because the hop is the part that goes wrong: an entity touching more thanGRAPH_MAX_ENTITY_DEGREE(40)
others is a hub — a namespace, a project every fact mentions — and is neither expanded from nor walked into, and a
query admits at mostGRAPH_NEIGHBOUR_CAP(25) neighbours and 40 facts. Measured on a 806-fact store: one generic
token resolved to an entity of degree 57 and an unbounded expansion admitted 110 weakly related facts into the
candidate pool.
A discovery never displaces a text hit insidesearch(): it fills the tail of the result list where the text side
left room, and only at half the weakest shown text hit. Withrerank: onthememory_searchtool and prefetch
rescore the whole pool afterwards, so a discovery that passesrelevant()can be ordered above a text hit and take
an injected slot; that is the rescorer's ordering, and docs/memory-graph.md says so. Measured on a live 807-fact
store, discoveries score ~0.05 against text hits of 0.17-0.42, so a ratio against the top hit never fired (0 of 15
realistic queries) and interleaving at slot 4 with the ratio removed pushed genuine hits down for filler. The
feature is live where recall is sparse — the case it exists for — and silent where the words already answer.
The reads are bounded too: at most 25 entity ids seed the walk, and the facts behind the neighbours — whose number
neighbour degree does not bound — are read in batches of 500, stopping at the 40-fact budget, so a store built
before SQLite 3.32 (999 bound parameters) does not fail a search. A reader that never migrated (a subagent, a cron
withcron_writes: false) has no entity_relations and simply gets no discoveries instead of an error.
Edges follow the entity index: a merge repoints them (_repoint_relations), a delete drops them, and forgetting a
fact removes the edge it implied unless another current fact still links the pair (trg_fact_entities_unedge). A
store written before the table existed gets a backfill fromfact_entitieson its first writer open, keyed on
meta.graph_backfilledrather than on an empty table.prune_graph()repairs edges no current fact implies;
rebuild_graph()recreates all of them. The graph schema is created by a writer's_migrate()only: a read-only
open (migrate=False) creates no graph schema (it still writes the WAL setup and the base schema, as before).
Off by default (graph_expansion: false), like the other experiments, until discoveries are measured reaching the
injected context on a real store. The first open of a store indexed under older entity rules rebuilds the entity
index (upstream behaviour,meta.entity_rules), and the graph follows it.
Holographic 0.26.6
Fixed
- Windows: a Hermes process no longer gives ~225 MB of commit charge to numpy's OpenBLAS once the plugin recalls.
OpenBLAS commits a buffer of ~32 MB per CPU thread when its DLL loads; Hermes core does not import numpy, the
plugin does on the first recall, and its math needs one thread. The plugin now setsOPENBLAS_NUM_THREADS=1for
that import only and removes it after, so the commands Hermes runs in child processes keep a numpy as before.
Measured on an 8-thread laptop: 243 MB of commit charge after the first recall, 18 MB now; the working set is the
same. Left alone: numpy loaded before the plugin, a thread count the person set (OPENBLAS_NUM_THREADS,
GOTO_NUM_THREADS,OMP_NUM_THREADS,OPENBLAS_DEFAULT_NUM_THREADS), and Linux and macOS, where the buffers cost
no resident memory. A child process started during that import (a fraction of a second) inherits the variable.
Holographic 0.26.5
Fixed
Found in a review of 0.26.3–0.26.4 (Claude Fable 5.1); with the presence gate off and no prefetch_max_tokens, only
the third item applies.
- Presence gate: a message longer than recall reads (4000 characters, a pasted log) with runs of spaces or newlines
no longer loses its turn. The prefetch reduced it after cutting, the sync before, so the two never matched and an
interrupted turn before it left the views a turn behind. - Presence gate: a session pushed out of the bounded records forgets its turns' messages with its turn count; a
message repeated after it came back could name a turn number from before. - With the journal on, a delivery that could not write its attempt drops only the receipts older than the last
compaction. Without a compaction between, the earlier block and its attempt line are still in context, and the
fact's next refresh no longer records an attempt again. - With the journal on and
prefetch_max_tokensset, a block whose new facts do not fit is fitted again as the
refresh it has become, without room for an attempt line, so a fact that fits is not dropped for it. docs/presence-gate.md: the limits of finding a sync's turn by its message.
Holographic 0.26.4
Fixed
Found in an Antigravity review of 0.26.1–0.26.3; with the presence gate off (the default) nothing changes except the
last two items.
- Presence gate: two late syncs in a row each keep their own turn. The first used to record the prefetch's turn, so
the second view was stamped with a later turn than its own and grace did not hold a newer injection. A sync now
finds its turn by its message (the prefetch of that turn saw the same one) and otherwise counts on from the
previous view, so an interrupted turn no longer leaves the views one turn behind. - Presence gate: a session that synced views, never recalled a fact and then compacted kept its generation,
compaction and summary records for the life of the process. They now live only while the session is tracked. - With the journal on and
prefetch_max_tokensset, a refresh (which carries no attempt line) no longer keeps room
for one, so a fact that fits is not dropped. tools/duplicate_count.py: a recall suffix without a recorded-on date ((this chat),(Alice · another chat))
no longer splits one revision.
Holographic 0.26.3
Fixed
- With the task-outcome journal and the presence gate on, a fact whose re-delivery after a compaction could not write
its attempt (a locked journal) kept the receipt of the block before the compaction, so every later delivery counted
as a refresh and the fact never recorded an attempt again. A delivery that needed an attempt and got none now drops
that receipt; the next delivery records its own (review of #79).
Holographic 0.26.2
Changed
- A prefetch that only refreshes facts this session already injected with a recorded attempt (the repeat window
expired, and no compaction or view of the context showed them leaving) records no memory attempt and carries no
Memory attempt:line. The earlier block and its attempt line are still what the model reads; with the core's
window dedupe (NousResearch/hermes-agent#148069) the repeated bullets are dropped, and the new attempt line was the
only thing left, keeping a heading-only block alive every few turns. Any new fact in the block, or a fact whose
earlier attempt was never written, still records an attempt. If the earlier block never reached the model (a
prefetch the core timed out), the refresh arrives without an attempt id; the plugin cannot tell.
Holographic 0.26.1
Fixed (experimental presence gate, off by default)
Found in a review of 0.26.0 (Codex and Antigravity); the default path is unchanged.
- A sync the core queued before a compaction no longer publishes the old view as current: the compaction credits
the summary row it was handed to the generation it ends. - A fact the compaction removed is recalled again once a view of the new context no longer shows it, instead of
waiting out the repeat window. - A compaction that fails after the hook no longer switches the gate off for the rest of the session: a view that
still starts with the transcript the hook was handed, grown by the turn, is taken as current. - A sync that starts after a later turn's prefetch keeps its own turn, so that prefetch's injections count as newer
than the view. - A
fact_storehistoryresult counts as visibility of the fact's current version. - Bounded the sync-origin record (it grew by one key per session that never recalled).
tools/duplicate_count.py: a cut line is recognised by the plugin's real rendering (300 characters ending in an
ellipsis), and(by meaning)/(inferred)no longer split one revision;tools/presence_delta.pyruns from a
checkout of any folder name.
Holographic 0.26.0
Added (experimental)
- Presence gate for prefetch (
prefetch_presence_gate, off by default; #72, issue #59). With it on, a fact this
provider injected is held back past the six-turn repeat window only while the model can verifiably still see
it: its exact emitted line inside a<memory-context>block the core fenced, or its(fact_id, content)
in afact_storeresult. Anything unclear falls back to the window. Off, recall is unchanged. Design and limits:
docs/presence-gate.md; duplicate countertools/duplicate_count.py. Tests:tests/test_holographic_presence_gate.py.
Fixed
changesays which keywords it dropped, andcorrectwrites the keywords it is given (#73, KostinPavel). A
changewithoutkeywordsdrops the old text's keywords (they describe the old text) but said nothing, so the
agent could not know to give the new text its own; the result now carrieskeywords_clearedwith those keywords
and how to set new ones. Acorrectwithkeywordsdropped them without a word: they now replace the fact's,
for a new text or a new start date alone, through the same write screen as onaddandchange, and the result
reports them asadd/change/updatedo; a correction that changes nothing writes none and says so
(keywords_not_set). The tool description and the prompt now say both. Tests:
tests/test_holographic_keywords_change_correct.py.- Tool results keep the fact's own words: raw Unicode instead of
\uXXXXescapes. All 33json.dumpscalls
in the tool handlers defaulted toensure_ascii=True, so a Cyrillic fact reached the model in every result
transliterated as escape sequences («Ловушка» as\u041b\u043e\u0432\u0443\u0448\u043a\u0430in enrich/history
outputs; found while discussing recall forms in #59). The handlers now serialize through_dumps
(ensure_ascii=False); network payloads (embeddings,rerank), stored digests (experience,history) and
the memory-map export are untouched, and ASCII-only results stay byte-identical. A payload that cannot be
encoded to UTF-8 (a lone surrogate, which can come from tool arguments) keeps the escaped form:_dumpsfalls
back, since the tool executor logs the result before anything else reads it. Expected side effect: the core's
size limits count characters, so Cyrillic results hit truncation or spillover less often. Tests:
tests/test_holographic_ensure_ascii.py.
Holographic 0.25.3
Fixed
- Starting the provider no longer imports numpy; the first search that scores by structure does. PR #45 moved
the module-level imports to first use, but the first use still came at start: opening the store asked
has_numpy()(which imports) for its vector checks, the retriever asked it to settle its weights, and
system_prompt_block()asked it for the "numpy is not installed" line. Nowholographic.numpy_installed()
answers those start-time questions without importing (the import's own answer once one was tried, else whether
Python can find the package); the retriever settles its weights on its first search; the store's vector repair
imports numpy only when there is a vector to write, and leaves the store as it was when that import fails (a
changedhrr_dimkeeps the old size on record). On a store that needs no repair, the usual open, a provider that
only starts (a Windows Kanban worker, NousResearch/hermes-agent#145156) leaves numpy unloaded. Still loading it at
open: a store whose vectors need re-encoding (facts added without numpy, anotherhrr_dim, new entity rules), and
search by meaning when on (the endpoint check). That alone does not stop the hang there: the first recall still
loads numpy, and the fix is in core (NousResearch/hermes-agent#144287, #146471). Tests:
tests/test_holographic_numpy_not_at_startup.py. - Search by meaning on no longer imports numpy at start either, and a broken numpy no longer takes memory down. The
provider's constructor askedEmbedder.unavailable(), which imported numpy; it now asks
holographic.numpy_findable()(no import), the import moves toEmbedder.usable()before a vector is needed, and a
recall checks that the store holds a vector of the model before that. numpy that Python finds but that fails to import
is treated as missing whatever it raises: a DLL that does not load comes out as ImportError, but numpy's own BLAS
sanity check at import raises RuntimeError, and its Windows wheel's DLL setup can raise OSError. Either used to raise
from the provider's constructor with search by meaning on; otherwise every write failed (fact_store add failed (OSError)), and on 0.25.2 memory was off for the session as "cannot open … if the file is damaged". Now recall goes
on by words, the log saysnumpy is installed but did not load (…)once per process, andstatsgives the reason
(structural_recall: off (numpy did not load: …), search by meaningunavailable: numpy did not load (…)). Tests:
tests/test_holographic_numpy_broken_or_embeddings.py. - A Latin word or identifier written into Chinese text is found by itself (NousResearch/hermes-agent#83593).
FTS5 indexes用户用keyring保存密钥as one token, letters and characters together, and the substring path ran only
for a query with Chinese in it, sokeyring,watchdog,ppid,HERMES_PARENT_PIDorweb_serveralone found
nothing while密钥did, and an English question about them recalled nothing before the turn. Search now also looks
for the query's ASCII words in the facts where a letter or digit touches Chinese, Japanese, Korean or Thai script,
as a whole word (keyis not inkeyring,pidnot inppid), and the recall gate counts those facts as holding
the word. The list of such facts is kept until a fact's text may have changed (a write other than the recall counts,
or a commit by another process), so a store without them reads nothing more per message: on 3,000 English facts
search and prefetch take the same time as before. On the multilingual quality set (450 questions in 18 languages) two
answers changed, both toward the other languages:HR24 Copilot能做什么ranks the fact that ends inHR24higher,
andhr-vpsのメモリはどれくらい?also recalls the otherhr-vpsfact, as the English, Russian and Chinese questions
do. Codex review: a letter of another script next to the word is no edge (яkeyring保存holds nokeyring), and
a fact the index already finds by the word (keyring保存, a prefix; the word also written apart) counts once in
its frequency (Codex review on GitHub found the same:keyring保存密钥fell out of the recall for "keyring watchdog
backend"). Codex review on GitHub: the list of such facts is built in linear time (a 30,000-character run with no
letter after it took 16 s). Not covered: other forms of the word (keyrings). Tests:tests/test_holographic_latin_in_cjk.py. auto_extractno longer stores a reply to the assistant as a preference (NousResearch/hermes-agent#22907).
"I like that, sounds good", "ok, I love it", "I prefer this one, let's go with it" and «ок, я люблю это» matched
"I like / love / prefer …", but what they like is whatever the assistant just said: stored, such a fact says nothing
on its own. A sentence whose object is "it", or this / that / these / those with at most one word, and whose clause
ends there (a degree word such as "better" or "a lot" may follow) is skipped, unless the rest of the sentence states
a preference of its own ("I like that, I always use tabs" is stored). Kept: "I like it when answers are short", "I
like that you ask before deleting files", "I use this laptop for work", "them" and «его» (more often people than a
proposal: "I love them, my kids are my everything"), and "I always check git status before committing" — a habit
said as one, which is what the "always / never / usually" pattern is there for. Measured: of the 91 sentences of the
test suite and benchmark results the patterns take, none of the earlier ones is skipped; of 36 preferences written
for the check, none. Codex review: the reply also ends at a dash or an emoji ("I like that 👍", "I love it — sounds
good"), and not at a dot inside a name ("I use this Node.js library for all my projects." is stored). Codex
review on GitHub: nor at a colon ("I prefer this format: JSON with two-space indentation." is stored). Tests:
tests/test_holographic_extract_reactions.py.- A prompt to an assistant is not stored as a fact, and one already stored is not recalled
(NousResearch/hermes-agent#66380). A mail bridge stored each inbound email as "You are Henry, … You received an email
at the AgentMail inbox and should handle it like a normal Henry request. EMAIL METADATA …", and recall brought every
one back, framed as stored data, as a live task in each later session for weeks. A text that opens by telling its
reader who it is ("You are …", «Ты — …», «Вы — …») and hands it a task it received (an email, a message or a task,
anEMAIL METADATAblock) is refused byadd,change,correct,update,revert,auto_extractand the mirror
of built-in memory, with a reason that says to store what is true instead;auto_extractskips such a message whole.
One written before is left out of the recall before a turn (logged once), andsearchandliststill show it, to
change or forget. Either half alone is no prompt: "You are allowed to push to main", "You are Mia, the family
assistant", "You received a parcel from DHL" are stored as before. English and Russian. Codex review: a standing
rule ("when I forward a message, you should handle it …") is no task received, so "should handle it" alone no
longer counts; «Вы — … Вам пришло письмо» does; the sentences of a pasted prompt are not extracted one by one; and
revertdoes not make such an earlier version current again. Measured: none of the 7,065 strings of the test suite
and benchmark results is refused; of 15 second-person rules and facts written for the check, none. A separate check
after_fact_problem, whose decisions stay 0.22.0's (tests/test_holographic_screen_speed.py). Tests:
tests/test_holographic_role_prompt.py. - The agent's report of its own steps is not stored as a fact (NousResearch/hermes-agent#117790). Besides
tool-call JSON (refused since 0.3.0), stores filled with "I just read the file and updated the config as requested."
A first-person action ("I read", "I've updated", "I ran", "I checked" …) together with words saying it was done for
the one the agent works for ("as requested", "as you asked", "per your request") is refused with a reason that says
to store what is now true instead. Both, so a person's own past ("I moved to Bangkok", "I switched to Postgres last
week", "I updated my passport in March") and a request of someone else ("as requested by the course") are stored as
before; a report without those words is still stored (the tool's description asks not to store task progress).
Codex review: a text the person said (originsaidorconfirmed, andauto_extract) is not screened so, since
there "I" is the person ("I updated my passport in March as you asked."); "just now" and "as requested by you"
with two spaces no longer slip through. English only. Measured: none of the 7,065 strings of the test suite and benchmark results is refused; of 16
first-person facts and 36 preferences written for the check, none. Tests:tests/test_holographic_narration.py.
Documented
auto_extractruns after each turn. README and the setup description said "at session end and before context
compression"; extraction also runs after every completed turn (sync_turn), and inscope: useronly then.