memory-slices

Memory slices — find the lesson + its recurrence, keep the record honest and bounded

A memory slice is a node's small, capped set of standing lessons: what this node has gotten wrong, the ROOT why, and the durable prevention — each lesson carrying a cross-run recurrence count and a [[okf-slice]] cross-reference to the code it concerns. It is a distillation, never a run transcript or a changelog. The lessons live in template-scoped memory.md — one per node (.piflow/<wf>/template/nodes/<id>/memory.md) plus one system-level (.piflow/<wf>/template/memory.md); the design/rationale is docs/research/memory/piflow-memory-v1.md §5a (the two legs) + piflow-memory-v1.5.md §6 (the MEMORIZE step in the overlord loop) — cite it, don't restate it. This skill is the OPERATIONAL procedure for the two things you do with memory.

The two legs, one pattern. okf-slices (Leg B) answers where the code is / how it works; memory-slices (Leg A) answers what went wrong here, why, and how often. The optimizer uses BOTH: memory-slices tells triage "this node has failed this way N times → SKILL, not LAPSE, and here's the lesson"; the lesson's [[okf-slice]] link hands the fixer the code map. Substrate is shared with Leg B: git (the deep, unbounded session log — skillsys/flowCommit commits) + the run traces (runs/<id>/, the raw material MEMORIZE distills) + codegraph via the OKF cross-ref. This "the commit message IS the record" move is the 2026 git-native-memory pattern — Lore (arXiv 2603.15566) restructures commit messages via git-trailers into decision records (constraints / rejected-alternatives / directives / verification); GitOfThoughts (arXiv 2606.14470) records scores as git-notes and merges a memory branch, and its finding that recalled memory helps only when a new case is >0.8 similar to a past one is exactly why we key on same-signature RECURRENCE; Projectmem (arXiv 2606.12329) is a grep-able append-only event log whose "you tried this before — it failed" judgment layer IS our recurrence signal, built. Do NOT confuse this template-scoped optimizer memory with two neighbours: (a) the run-scoped, forward-INJECTED MEMORY.<node>.md some products write for in-run coordination (opposite semantics — never touch it here); (b) the memoryDir session hub (~/.claude/.../memory) that OKF slices already read for their code lessons-cluster. This skill owns only the template memory.md.

Why this exists: without a bounded, cross-run record, every optimizer round re-discovers the same defect and can't tell a one-off slip from a real skill gap — so it either never edits (misses SKILL defects) or edits on a single slip (rots the corpus). Memory is only useful if (a) you can FIND the right lesson + its recurrence and (b) it is FRESH and SMALL. This skill is both halves.

---

MODE A — FIND the lesson + recurrence for a node (the reader)

Use when triage is about to bucket a defect, or a fixer is about to edit a node, and you need the node's standing record instead of re-deriving it. This is the reader that fills the triage.ts SKILL-vs-LAPSE deferral. Procedure (stop at the first step that resolves):

1. Normalize the query to concrete keys: the NODE id, and the failure SYMPTOM (a short signature — the failing check id, the error class, the artifact field, or a one-line description).
2. Read the node's memory.md (.piflow/<wf>/template/nodes/<node>/memory.md) and match the symptom against its lessons. RANK by where the match lands — a lesson whose Symptom/signature matches OWNS the query; a bare prose mention is a WEAK match. If the defect's fix-surface is cross-node (a hand-off/contract), also read the system memory.md.
3. Compute RECURRENCE deterministically — never guess it. Count prior occurrences of this signature across runs: the lesson's own recurrence: counter, corroborated by the run trail (grep the signature across runs/*/ reports/traces) and git (git log --grep=<signature> over skillsys/flowCommit commits). The COUNT is the signal; the prose is the explanation. Grep accuracy depends on a DISCIPLINED, controlled-vocabulary signature — CommitDistill (arXiv 2605.18284) shows an un-disciplined git mine needs a calibrated abstention threshold to stay correct; prefer stamping the count as a git-note on the landing commit (GitOfThoughts) so it is git-auditable and mergeable, not only in the memory.md frontmatter. If nothing matches → recurrence 0 (first occurrence).
4. Validate the lesson before you trust it (just-in-time; a stale lesson is worse than none): a lesson pinned to code via [[okf-slice]] is only current if that slice is. Run the OKF gate for the linked slice — cd .agents/okf/topics && node _generate.mjs --check <okf-key> — and read the signal (see okf-slices MODE A for HEALTH vs DRIFT). HEALTH: on the linked slice ⇒ the code the lesson describes MOVED: the lesson's prevention may no longer apply → flag it code-shifted, don't hand it over as settled. ok/DRIFT ⇒ the lesson's code anchor still holds.
5. Return the SIGNAL, not a dump. Hand triage: the matching lesson (Symptom → Root → Prevention), the recurrence COUNT, the linked [[okf-slice]], and the freshness verdict. The count drives the bucket — recurrence 0 → LAPSE (default-when-unsure, protect the corpus); recurrence ≥ threshold with a prevention the node already had → SKILL (the prose was right, executor slipped again → routing/tier, not a prose edit); recurrence ≥ threshold with NO prevention yet → SKILL (author the missing rule). A code-shifted lesson routes toward FUNCTIONALITY/ARCH, not SKILL.

Recurrence ≠ certainty. A count says the symptom repeated, NOT that the same ROOT caused it. Before you let a high count force SKILL, confirm the roots match (same trace signature, same fix-surface) — two different bugs with one symptom are two LAPSEs, not one SKILL. When the roots diverge, report ambiguous and default to LAPSE.

FIND output shape (what you return to triage/the fixer): { node, lesson? (symptom/root/prevention), recurrence: N, okfSlice?, freshness: fresh | code-shifted | none, suggestedBucket, confidence }. If no lesson, say so plainly (recurrence 0) — never invent one.

FIND bar (must hold): you cited a REAL lesson (or honestly reported recurrence 0); the recurrence count was COMPUTED from traces/git, not asserted; you confirmed roots match before calling recurrence a SKILL signal (else ambiguous→LAPSE); you ran --check on the linked OKF slice and reported freshness; you returned the signal triage needs (bucket + evidence), not the whole file; you did NOT present a stale/code-shifted or invented lesson as settled.

---

MODE B — MAINTAIN the memory set (MEMORIZE + keep it bounded)

Use when the optimizer's MEMORIZE step runs, after a fix lands, when the cap is hit, or when asked "is this stale / where should this lesson live." Memory has two surfaces, split by write-authority × freshness — keep them distinct:

• git = the deep, unbounded session log. Every landed edit is a skillsys(<node>)/flowCommit commit whose message IS the record (Symptom → Root → Fix → Prevention). NEVER hand-keep a changelog file — that is the #1 memory-rot; git is the generated log. git is QUERIED (log/grep/blame), never loaded wholesale into context.
• memory.md = bounded STANDING STATE. Lessons + why + current invariants + open threads. NOT diffs, NOT a log. Capped (~40 lines, top-loaded; lowest-value truncates first). Optimizer-facing; never injected into the node.

The write authority + the update rule. Only the optimizer's MEMORIZE step (and a human) writes memory.md — disjoint from the node it describes. The rule is ACE incremental-delta, never full-rewrite:

• append a new lesson (with recurrence: 1, a [[okf-slice]] link, and the distilled Symptom/Root/Prevention);
• update an existing lesson (increment recurrence, sharpen the prevention) when the same signature recurs;
• retire a lesson that is superseded (its fix graduated to code/git) or stale (its [[okf-slice]] went HEALTH and the prevention no longer applies). This append / update / retire shape is ExpeL's insight-pool operations (ADD / EDIT / UPVOTE / DOWNVOTE, arXiv 2308.10144) — our recurrence counter IS ExpeL's vote weight — and Mem0's ADD/UPDATE/DELETE with newer-wins (arXiv 2504.19413). Zep/Graphiti invalidates rather than deletes (arXiv 2501.13956), but git is already our invalidation ledger, so memory.md can hard-remove a retired lesson without losing history. Do NOT LLM-rewrite the whole file into a fresh summary — full-rewrite consolidation causes measured context collapse (ACE, arXiv 2510.04618; A-MEM, arXiv 2502.12110, DOES let a new note evolve/rewrite existing linked notes, but that optimizes retrieval recall for an unbounded, runtime-INJECTED store — the wrong move for our bounded, optimizer-only, read-whole corpus; do not read it as license to rewrite lessons). When the cap is exceeded, compaction is a SEPARATE, out-of-band pass that RETIRES discrete lowest-value entries (by age × non-recurrence × graduated-to-git), never a re-summarization. The ~40-line cap and the retire-metric weights are TUNABLE defaults, not laws — the eviction surveys (Du et al. arXiv 2505.00675; the LLM-agent-memory surveys) settle the AXIS (age × importance × recurrence), not the constant.

Where a lesson lives (the assignment rule) — by the four-way triage bucket, i.e. by blast radius:

Cadences:

• Per-round (the MEMORIZE step of the overlord loop). After SCORE→TRIAGE→FIX→GATE, distill the round's confirmed defect into ONE delta (append/update/retire per above) on the owning memory.md, link its [[okf-slice]], and stamp the recurrence. This is the write; it is out-of-band, never an in-DAG node.
• Compaction (when over cap, out-of-band). Retire discrete entries by the metric above; never re-summarize.
• Post-merge freshness (advisory). When code moves, a lesson can go stale. Re-run the linked slices' okf --check; any lesson whose [[okf-slice]] is HEALTH-flagged → mark code-shifted for the next MEMORIZE to re-validate or retire. The memory drift gate is not separate machinery — it RIDES the OKF gate through the link.

To WRITE a lesson (the delta shape): a lesson block = ### <symptom signature> · recurrence: N · [[okf-slice-key]] · Root: one line · Prevention: the durable rule (positive recipe for SKILL, prohibition+rationalization for a discipline lapse — match the form to the failure type, see agentic-prompt-design §0b). Keep it pointer + semantics: link the slice, cite the trace path:line, do NOT paste the code or the transcript.

---

Invariants (the laws — do not violate)

• Optimizer-facing, never injected. memory.md is read/written by the out-of-band optimizer; it is NEVER put into the worker node's own runtime prompt (a node that sees its failure history only hesitates). Keep memory.md out of any directory a worker node's tools sweep.
• Distillation, never transcript. A lesson is Symptom/Root/Prevention + pointers; it does not duplicate the run log or the code. git holds the history; memory.md holds the standing state.
• Deterministic-first; the model only distills. Recurrence is COUNTED (traces + git); the LLM writes the lesson prose. Never let a model assert a recurrence count.
• Incremental-delta, never full-rewrite. Append/update/retire discrete lessons; compaction retires, it does not re-summarize. Bounded by the cap.
• Validate after retrieval (JIT), don't front-load. Pull the lesson when needed and ride the linked slice's --check before trusting it; a stale/code-shifted lesson is actively harmful.
• git is the log; never hand-keep a changelog. The commit graph is the record; memory.md is not a diff feed.

Self-check before returning

• FIND: Did I cite a real lesson (or say recurrence 0), COMPUTE recurrence from traces/git (not assert it), confirm the roots match before calling it SKILL (else ambiguous→LAPSE), ride the linked slice's --check and report freshness, and return the triage signal (bucket + evidence) not a dump?
• MAINTAIN: Did I write ONE incremental delta (append/update/retire) — never a full rewrite? Did I place the lesson by its bucket (node vs system)? Did I keep it distillation + pointers (no transcript/code)? Did I leave git as the log (no changelog file)? On compaction, did I retire discrete entries rather than re-summarize?

Pointers

• Design + rationale: docs/research/memory/piflow-memory-v1.md §5a (two legs) · piflow-memory-v1.5.md §6 (the MEMORIZE step + the overlord loop) · §3 (the four-way triage the assignment rule mirrors).
• Leg-B twin (the pattern to match) + the shared cross-ref: .claude/skills/okf-slices/SKILL.md; a lesson's [[okf-slice-key]] is a key under .agents/okf/topics/.
• The reader stub this fills: packages/core/src/optimize/triage.ts:14 (SKILL ← cross-run recurrence, DEFERRED).
• The writers already seeded: packages/core/src/memory/{skeleton,seed}.ts; piflowctl memory scaffold <template>.
• SOTA grounding + what's settled vs open: docs/research/memory/memory-substrate-sota-2026-06-30.md (git-native substrate — Lore/GitOfThoughts/Projectmem/CommitDistill · the substrate comparison · the ExpeL/ACE/Mem0 update-rule + eviction SOTA · G-Memory/Wu-&-Shu MAS scoping). RESOLVED there: flat-markdown-on-git is the right substrate (vector/graph/SQL is overkill for a bounded, read-whole, optimizer-written corpus; Mastra semantic-recall targets an unbounded conversational corpus we do not have); the memoryDir hub stays separate. STILL OPEN (tunable, not law): the ~40-line cap constant + retire-metric weights, and the optional per-vertical tier.
• Promotion path (Stage 2, mirroring okf): port FIND/RECURRENCE into a deterministic piflowctl memory find|check verb (a deriveRecurrence() alongside the generator's deriveLessons()), then add this skill to the piflowctl skills install bundle.