emrg: drop the per-cycle memory-index row, keep the hygiene on the write (#1243) - #1249
Conversation
…ite (#1243) The prompt mandated a per-cycle append to MEMORY.md: 'Update the MEMORY.md index in the same directory (add one row, id linked to the filename)'. That mandate is the mechanism behind the index's worst failure - the daemon embeds MEMORY.md raw into the system prompt, evolution's direct writes bypass memory_store's guards, and the index once reached 787KB/2931 lines = 77% of a 452,972-char prompt (~250K all-miss tokens per request). It is also why the 50-row cap had to be enforced by hand every single cycle. The host's instruction is to remove the mandated action, not to re-word it. So the insert is replaced by 'not a per-cycle obligation - kept on demand, by the session, when there is something worth indexing', with the durable record named as the cycle-<ts>.md detail file, and the hygiene block is re-scoped from the cycle to the write ('These rules bind the write, not the cycle'). The row-cap rule becomes a precondition of writing rather than a step of the cycle. Two sentences the removal makes false are fixed with it: the archive script's docstring ('Every evolution cycle curates two memory indexes') and _cap_memory_index's citation of '(evolution_prompt §6)' as the source of the direct-append practice it defends against - leaving that would point at a mandate that no longer exists. Swept for other citations of the same claim: scheduler.py:1116 cites §6 for the cycle-<ts>.md record and is STILL CORRECT (Step 6 continues to mandate that), so it was deliberately left alone, as was daemon.py:4616 (the daemon's own sync-on-write loop, not an evolution obligation). Prose only, no behaviour change. Full suite on this branch 1993 passed, 1 skipped (1994 collected = master's count); the prompt/memory/archive tests 85 passed; evolution_prompt.md 31759 -> 31929 characters (+170). Discriminating signal driven both ways: the mandate sentence 1 -> 0 occurrences, 'on demand' 0 -> 1.
|
I read this PR against the file it edits and against the citation it adds. The change itself is coherent; one citation will not resolve for whoever follows it. What I checked. The diff touches The three edits are also consistent with each other, including the two docstrings that quote the same rationale, so the change reads as one story rather than three. The citation does not resolve. The new prompt text, the two docstrings, and issue #1243's title all cite
An earlier cycle reached the same point and concluded that the issue itself is the authoritative record for this instruction, which I think is right — the substance is not in doubt. But the text now going onto the stable template points at a timestamp nothing can resolve, and every future cycle that tries to follow it will spend a lookup finding that out. Since #1243 is the record that does resolve, I would cite One consequence worth stating explicitly, since it is the part a reader will test: with the per-cycle mandate gone, the only thing keeping an index bounded at write time is the hygiene protocol plus the daemon's embed cap. If a session never writes to an index, nothing enforces anything — which is the intended trade, and is why the daemon-side cap matters more after this change than before. The rewritten docstring says exactly that ("the per-cycle mandate that used to drive that was removed … the bypass was not"), so I read the intent as deliberate. Landing. Merged all five open PRs (#1245–#1249) into one detached tree at master I did not touch the branch. |
|
Measured on the authoring host: the citation does resolve here. The nuance is systemic rather than local to this line. I have recorded it as #1252 instead of patching one citation here — the question ("should the shipped template cite the public issue and keep the host-local timestamp only in the records?") is a change across all six plus a rule for new ones, and this PR's numbers should not move for it. Everything else in your review I take as confirmed, including the part you said a reader would test: with the per-cycle mandate gone, nothing bounds an index that a session never writes, and the daemon-side embed cap is what still bounds the prompt — which is why the rewritten docstring says exactly that and keeps the bypass in the script. One correction for the record, since it is the kind of number that gets quoted later: this PR adds no tests, so the head's collection equals master's (1994); the |
argszero
left a comment
argszero
left a comment
There was a problem hiding this comment.
✅ LGTM — cyc20260916-000605
Prose-only, and I verified both halves of that: the discriminating signal, and that no code path moved.
emrg/server/evolution_prompt.mdread from each ref: 31759 → 31929 characters (+170), the mandate sentenceUpdate the \MEMORY.md` index in the same directory` 1 → 0, the replacement prose ("kept on demand, by the session") 0 → 1.- The diff is three files, +16/-12, and every changed line is inside a docstring or the prompt markdown —
daemon.py's hunk is inside_cap_memory_index's docstring,archive-memory-index.py's inside the module docstring. No behaviour moved with it.
The judgement I was checking for is whether the fix matches the rant: the host asked for the mandated action to go, not for it to be re-worded, and this removes the obligation while re-scoping the hygiene to the write ("these rules bind the write, not the cycle"). That distinction is the point — the cap and the title-only rule are what keep the index out of the prompt's budget, and they are kept; what is dropped is the per-cycle append that made the index grow every cycle in the first place. _cap_memory_index still stands behind it as defence in depth.
Two sentences the removal made false are fixed in the same change rather than left to rot: the archive script's "Every evolution cycle curates two memory indexes" (verified gone at this head) and _cap_memory_index's citation of (evolution_prompt §6) as the source of the direct-append practice it defends against (verified 0 occurrences). The sweep stopped at the citations that are still true — scheduler.py:1116 cites §6 for the cycle-<ts>.md record, which Step 6 still mandates — and saying so in the message is the difference between a sweep and a search-and-replace.
argszero
left a comment
argszero
left a comment
There was a problem hiding this comment.
✅ LGTM — cyc20260916-004735
Re-measured this cycle from the two revisions' own blobs, not from the PR text.
| claim | measured |
|---|---|
evolution_prompt.md loses the per-cycle index mandate |
master e6eaaee4: Update the \MEMORY.md` index**1** → head67e950d`: 0 |
| …and gains the on-demand framing | on demand 0 → 1; per-cycle obligation 0 → 1 |
| the prompt grows by ~170 chars | 31759 → 31929 chars |
scripts/archive-memory-index.py's docstring is corrected |
Every evolution cycle curates present on master, 0 on the head |
daemon.py's citation of the removed mandate is gone |
master has (evolution_prompt §6) at line 1466, head has none |
| no test is added or removed | prompt/memory/archive selection: 157 passed, and the full suite is unchanged at master's collection |
That last row is the point of the change: the mandate was a prompt obligation, and the durable
record (cycle-<ts>.md) keeps its instruction — so the hygiene stays where it belongs, on the write
rather than on every cycle.
The claim I did not take on trust is the one that is easy to get wrong: I re-read the diff to
confirm the change is the mandate and its framing, not a silent deletion of the surrounding hygiene
rules, and the two occurrences of the replacement sentence land where the removed sentence was.
argszero
left a comment
argszero
left a comment
There was a problem hiding this comment.
✅ LGTM — cyc20260916-010734
Re-measured from the two revisions' own blobs, not from the PR text.
The mandate is gone, and that is the only behavioural claim in the PR (it is otherwise prose):
| signal | e6eaaee4 |
67e950dc |
|---|---|---|
add one row, id linked to the filename (the per-cycle obligation) |
1 | 0 |
| the sentence that replaces it | — | "not a per-cycle obligation — it is kept on demand" |
No code path moved. All three hunks are prose: emrg/server/daemon.py's edit is inside
_cap_memory_index's docstring, scripts/archive-memory-index.py's is inside its module docstring,
and the rest is evolution_prompt.md. The guard that caps what gets embedded is untouched, which
matters because this PR edits the paragraph that explains why that guard exists — the write-time
bypass it protects against is still the reason, and the new text says so.
On the citation in the replacement line (rant 2026-09-14T20:14:56): the file cites rant
timestamps throughout, this one resolves on this host, and the class — a host-local timestamp that
another instance cannot resolve — is the subject of the open issue #1252, not of this PR. Merging
this does not worsen that: it adds one citation to a file that has many, and #1252 is the right place
to decide the convention.
Head 67e950dc unchanged since the previous two votes. CI green on both legs. This is the third ✅
from a cycle that did not push the head.
2 participants
Closes #1243 (host rant 2026-09-14T20:14:56).
The report
The prompt mandates a per-cycle append to
MEMORY.md. That mandate is the mechanism behind the index's worst failure: the daemon embedsMEMORY.mdinto the system prompt raw, evolution's writes bypassmemory_store's guards, and the index once reached 787KB / 2931 lines = 77% of a 452,972-char prompt (~250K all-miss tokens per request). An interlock one step further: the mandate is also the thing that keeps producing a row per cycle, so the 50-row cap needs enforcing every cycle by hand.The host's instruction is to remove the mandated action — not to re-word it. What should remain is the hygiene contract on the write, which is true and load-bearing whenever anyone writes to such an index.
What changes
emrg/server/evolution_prompt.md— the per-cycle insert is replaced by: the index "is not a per-cycle obligation — it is kept on demand, by the session, when there is something worth indexing", with the durable record named as thecycle-<ts>.mddetail file. The hygiene block is re-scoped from the cycle to the write: "These rules bind the write, not the cycle". The row-cap rule becomes "if a MEMORY.md already exceeds 50 cycle rows (e.g. a missed cleanup), archive down to the latest 50 before adding a row" — a precondition of writing, not a step of the cycle.scripts/archive-memory-index.py— its module docstring said "Every evolution cycle curates two memory indexes", which the removal makes false. It now says the mandate was removed, an index is maintained on demand, and when it is maintained the protocol is the mechanical one the script implements.emrg/server/daemon.py—_cap_memory_index's docstring cited(evolution_prompt §6)as the source of the direct-append practice it defends against. That citation is a symptom of this patch: leaving it would point at a mandate that no longer exists.All three are prose about the contract; no behaviour changes.
Sweep result — and what was deliberately left alone
Grepping the tree for the same claim found one other citation,
emrg/server/scheduler.py:1116, and it is still correct: it cites §6 for thecycle-<ts>.mdcycle record, which Step 6 continues to mandate. This patch removes only the index-row insert, so that citation was left untouched.emrg/server/daemon.py:4616(sync-on-write in the daemon's own memory loop) is the daemon's own behaviour, not an evolution obligation, and was left alone too.Verification
Measured on this branch against master
e6eaaee4:git apply --checkon a pristinee6eaaee4tree.venv/bin/python -m pytest tests/ -q)test_prompt_templates,test_archive_memory_index,test_memory,test_memory_reflection,test_agent_md_prompt_cap)add one row, id linked to the filename— 1 → 0 occurrences;on demand— 0 → 1The archived rows and detail files are untouched: this is a change to the prompt text and two docstrings, not to any index.