can four agents remember the same thing?

optmem's structural foundation is integrated. optchat is a separate harness experiment. the original cohort is still partial, so implementation progress and measured recall stay separate.

what actually landed in optmem

the old page stopped at the september 29 cohort. work continued, but the publication did not. this is the missing implementation update, frozen to the public revisions linked below.

first, credit where it belongs: Victor Taelin's OptMem is the upstream reference for the temporal-memory approach. Kevin directed the z0 integration and identified the missing publication. our adaptation and its evidence are separate from the original project.

z0intelligence integrated the event ledger, optmem-style temporal projection, and scoped memory contract through reconciliation pr #109 at b345aa2. the exact source is inspectable in the pinned reconciliation record and pinned temporal projection.

events.jsonl stays canonical. the tree is a rebuildable view over older history; recent events remain raw, sticky historical events can be expanded, and zoom() returns the original event payloads. rebuilding that view does not turn it into another authority.

optchat is a different layer

shared memory and a harness's compacted chat history are related, but they are not the same store. the new optchat work keeps those responsibilities separate:

slice	status on october 5	boundary
optmem foundation	integrated structural baseline	canonical event log plus derived temporal view; no semantic-quality qualification
omp optchat	open pr #126	append-only harness chat log and compactor; not a new shared memory store
hermes optchat	draft pr #127, stacked on #126	optional off / shadow / on adapter; default off, separate Hermes log root

the omp snapshot is 0828b77 in pr #126. the hermes adapter is fc31fb5 in pr #127. these are pinned implementation snapshots, not a claim that either open branch has reached the runtime default.

#126 reports focused optchat/retrieval-replay tests. #127 reports green core-unit and no new optchat, hermes-adapter, or memory failures relative to its base, while the full offline suite still has base failures. that is narrower than saying the whole stack is green. publishing this update does not merge those runtime prs or enable either adapter.

the next missing evidence is comparable answer support, provenance, contradictions, abstention, context size, and cold/warm latency against the same frozen questions. private chat logs and personal memory stores are not imported into this public report.

memory is not one thing

this kept biting us because we would say “memory worked” when we actually meant one of like four different things.

retrieval → model-visible injection → answer support → verification

those are not interchangeable. finding a row in an index is not the same as putting it in front of the model. putting it in front of the model is not the same as the answer using it. and even a supported answer is not automatically verified.

freeze the questions before touching the harnesses

i did not want dsh, hermes, omo, and omp each “passing” their own custom little memory demo. so the cohort is frozen first:exact-identifier, supersession, cross-harness, contradiction, missing-evidence, and minimal-context.

the contradiction case has to show both claims and pick no winner. the missing-evidence case has to say “i do not have it.” honestly those two are more interesting than the easy recall cases anyway.

omp finally gave us a number i actually care about

omp pr #107 runs the same six question ids through the common receipt schema and passes 16/16 focused tests. the fun part is the context size at turn 4:

native

320,075 b

+80,008 b growth

donor spill stub

2,847 b

+701 b growth

state packet

960 b

+0 b growth

reduction

99.70%

333.4× smaller than native

960 bytes versus 320,075 bytes is the kind of difference that changes architecture. but it only says we shaped context way better. it does not magically prove all four harnesses remember correctly.

agentsview also failed closed on that host. good. no hit was invented just to make the demo look green.

do not average pending work into a headline

this table is intentionally boring. boring is good here.

harness	status	what we actually have
omp	measured	common receipts, 16/16 tests, turn-4 context measurement
dsh	bounded	default-off agentsview stdio overlay; frozen model-visible proof still pending
hermes	pending	final comparable cohort receipts not imported yet
omo	pending	final comparable cohort receipts not imported yet

i am not turning one good omp row into “unified memory solved.” once all four lanes produce the same receipt shape against the same frozen cohort, then we can talk.

what this still does not prove

one omp lane does not prove four-harness equivalence;
smaller context does not prove better answers by itself;
we do not need one giant universal memory database for this to work;
retrieved text is evidence, not automatically truth;
there is no cross-harness aggregate until the comparable receipts exist.

the thing i actually want is simpler: same question, same evidence contract, different harness, same answer support. once that is boring, memory stops being a product-specific magic trick and starts becoming infrastructure.