solution-erp/.claude/agent-memory/memory-budget.json

{
  "_note": "Harness-9 (S70, 2026-06-17) memory budget. Caps SEEDED BY MEASUREMENT (scripts/measure-agent-memory.ps1), NOT imagined headroom. Budget-audit @session-start (session-start.md §2.1.2): if curate-to-fit is dropping important markers, BUMP the relevant cap rather than cut markers. Re-measure with the script; never hand-edit measured_bytes.",
  "seeded_date": "2026-06-18 (S71 re-measure after G1 curate — reviewer 36.7KB + investigator 29.8KB over-cap from S71 same-role race, curated L1->L2 back under auto-inject cap)",
  "last_sleep_at": "2026-06-18",
  "_last_sleep_at_note": "Harness-10b sleep-recovery (S72): timestamp lan cuoi chay /sleep-recovery-memory-l2. null = chua tung. session-start §2.1.2 + session-end §L.b(c) doc field nay -> INFORM goi-y nen L2 neu null hoac today-last_sleep_at>=7 ngay. Lead=single-writer (chi command sleep set field nay).",
  "tiers": {
    "l1_hot": {
      "file": "MEMORY.md",
      "injected": "auto (harness injects ~first 200 lines / 25KB on spawn)",
      "autoinject_cap_bytes": 25600,
      "soft_cap_bytes": 30720,
      "rule": "keep MEMORY.md < autoinject_cap so the WHOLE hot file injects; over soft_cap => curate L1->L2"
    },
    "l2_index": {
      "file": "archive/_INDEX.md",
      "injected": "read-on-demand map (inject the map, not the territory)",
      "cap_bytes": 20480,
      "seeded_from": "max measured _INDEX = cicd-monitor 16779B (+ ~22% headroom). cicd-monitor at ~82% = WATCH-agent (grows with each run); when it nears cap -> gist-of-index or split, do NOT drop markers",
      "pointer_style": "substring (git-sha / Run#NNN / unique-phrase keyed), fallback Ctrl-F. NO line-hints (additive appends shift lines)"
    },
    "l2_verbatim": {
      "file": "archive/<period>.md",
      "injected": "read-on-demand content (frozen, additive-only). NO inject cap",
      "rule": "NEVER rewrite existing bytes; curated L1 entries APPEND to end only"
    },
    "l2_gist": {
      "file": "archive/<period>.gist.md",
      "injected": "read-on-demand 4-field distill (distill-gen counter guards re-distill). NO inject cap",
      "rule": "coverage-diff GATE: every surprise/guard/file:line/root-cause/gotcha# in verbatim must survive in gist (or marked N/A)"
    }
  },
  "archive_gate": {
    "_note": "Harness-11 PART-A (S73, 2026-06-18) params for scripts/memory-archive-gate.ps1 (DRY-RUN planner, NO-API, grep+measure only). ADDITIVE — does not touch measured/tiers/last_sleep_at. A4 hysteresis = drain to BELOW low_watermark (not just to the line). A5 = never auto-drain below keep_floor newest entries (WARN instead). A6 = only PROPOSE archive after strike_threshold consecutive over-cap runs (stateless script persists strikes in .claude/agent-memory/.archive-strikes.json).",
    "autoinject_cap_bytes": 25600,
    "low_watermark_ratio": 0.85,
    "keep_floor_entries": 5,
    "strike_threshold": 2,
    "value_protect": {
      "_note": "Harness-15 B(b) value-gated archival (S81, 2026-06-20): keep_floor_entries protects NEWEST-n (recency/age axis); value_protect protects HIGH-VALUE entries regardless of age (value axis, orthogonal). Recurring-bug / anti-pattern / gotcha / root-cause entries STAY in L1-hot even when old -- archival cuts LOW-VALUE, NOT FIFO-by-date. This is mark RC-...10-29-11 (time/age=false-proxy) applied to the memory layer. Mechanism = CONVENTION (em-main judgement when condensing L1->L2); the patterns below are an advisory grep-hint for the DRY-RUN planner to FLAG protected entries, NOT an enforced auto-exclude (no overclaim: archive-gate stays DRY-RUN, em-main decides).",
      "patterns": ["gotcha #", "anti-pattern", "recurring", "lost-update", "race", "bai hoc", "lesson", "guard", "root-cause", "silent-fail"]
    }
  },
  "harness_floor": {
    "_note": "Harness-15 A1/A3 (S81, 2026-06-20): SAN-harness = fixed per-spawn cost (NOT tunable) = tool-schema + framing + own persona/role file + lead-pasted base-doc slice + task prompt. SEPARATE HOUSE (A3 anti-double-count): persona + lead-pasted-docs belong HERE (floor), NOT counted in token_governor.l1_always (which = own agent-memory + archive index + work-state block only). MEASURED-ESTIMATE not exact (H15 honest-note b): persona = directly measured bytes (.claude/agents/<name>.md 4.3KB-13.3KB => ~1.3K-4.0K tok via /3.3); tool-schema + framing = harness-injected (cannot byte-count locally), estimated comparable to AI_INFRA same-toolset-family (Read/Write/Edit/Bash/Grep/Glob/Skill/RAG).",
    "measured_token_estimate": 21000,
    "rounded_up_tokens": 30000,
    "components_note": "persona(measured 1.3K-4.0K, DIRECTLY via Get-ChildItem byte-count of .claude/agents/*.md, NOT via measure-agent-memory.ps1 which only sizes agent-memory tiers) + tool-schema(est ~5K) + framing(est ~2K) + lead-pasted base-slice(est ~5-9K; em-main injects CLAUDE.md/README-slice + task ctx) + prompt(est ~1K). README decision-tree corpus = 32704B but em-main pastes a SLICE, not whole. The ~21K sum is SE's OWN reasoning over SE's OWN agents (persona byte-measured + each harness-injected sub-component estimated independently); it lands NEAR AI_INFRA's ~21K because the toolset family is identical (Read/Write/Edit/Bash/Grep/Glob/Skill/RAG), NOT because borrowed. SE's governing cap = 30K (SE's own round-up), independent of AI_INFRA's figure."
  },
  "token_governor": {
    "_note": "Harness-15 A2/B(a)/B(e) (S81, 2026-06-20): SECOND governor (token) ORTHOGONAL to the BYTE governor (tiers/archive_gate above). Keep BOTH (B(e)) -- byte measures file-size-on-disk, token measures context-loaded; independent (an agent can exceed byte-cap yet load FEWER tokens; VN text ~3.0-3.5 byte/tok so byte/4 = upper bound => real headroom LARGER). Numbers = LEAD-AUTHORITY hard-cap derived from SE byte-caps x workload (multi-module ERP => heavier budget); NO AI may re-optimize them down (B(a)). Budget = MINIMUM-to-USE floor, NOT a ceiling to dribble against: FILL L1 with real work-state up to budget; under-fill ONLY when high-value content exhausted; NEVER garbage-stuff (token-saving = forgetting work).",
    "l1_always_tokens": 12000,
    "l1_always_note": "own agent-memory (MEMORY.md ~8K @ 25600B/3.3) + archive _INDEX map (~2K) + work-state block (~2K). Always-loaded.",
    "l2_ondemand_tokens": 6000,
    "l2_ondemand_note": "archive verbatim/gist sections + skill sections. Pulled per-need, NOT always-loaded => no context-rot when unused.",
    "l3_rag_tokens": 4000,
    "l3_rag_note": "RAG search_memory/search_code result per query. On-demand.",
    "headline_floor_plus_l1_tokens": 42000,
    "headline_note": "always-present per spawn = SAN(30K) + L1(12K). L2/L3 expand only on-demand (no always-cost)."
  },
  "measured": {
    "_note": "S81 2026-06-20 FRESH re-measure (scripts/measure-agent-memory.ps1) post-S80 curate. Supersedes stale S71-seeded values. byte-governor snapshot (l1_hot = file-size-on-disk); cross-check token_governor for the orthogonal token-thread.",
    "cicd-monitor":          { "l1_hot": 12227, "l2_verbatim": 194964, "l2_index": 17626, "l2_gist": 29737, "rollout": "done (re-curated S80: keep-floor-hit manual condense, ~3 huge run-records -> archive)" },
    "investigator-codebase": { "l1_hot": 10281, "l2_verbatim": 102446, "l2_index": 12308, "l2_gist": 27297, "rollout": "done (re-curated S80: 15 recon entries -> archive)" },
    "reviewer":              { "l1_hot": 10242, "l2_verbatim": 92316,  "l2_index": 10067, "l2_gist": 19114, "rollout": "done (re-curated S80: 13 war-story entries -> archive)" },
    "implementer-backend":   { "l1_hot": 19394, "l2_verbatim": 59233,  "l2_index": 10105, "l2_gist": 23079, "rollout": "done" },
    "frontend-designer":     { "l1_hot": 26129, "l2_verbatim": 0, "l2_index": 0, "l2_gist": 0, "rollout": "n/a (no archive); OVER autoinject-cap 25600 (under soft-cap 30720) -- WATCH strike-1, first-overflow needs _INDEX/gist build" },
    "test-specialist":       { "l1_hot": 27723, "l2_verbatim": 5546, "l2_index": 0, "l2_gist": 0, "rollout": "n/a (tiny archive, no _INDEX); OVER autoinject-cap 25600 (under soft-cap) -- WATCH strike-1; value-protect FLAG fired S81 (gotcha#/guard in move-set)" },
    "harvest-curator":       { "l1_hot": 18952, "l2_verbatim": 0, "l2_index": 0, "l2_gist": 0, "rollout": "n/a (no archive)" },
    "tooling-auditor":       { "l1_hot": 18431, "l2_verbatim": 0, "l2_index": 0, "l2_gist": 0, "rollout": "n/a (no archive)" },
    "implementer-frontend":  { "l1_hot": 17386, "l2_verbatim": 0, "l2_index": 0, "l2_gist": 0, "rollout": "n/a (empty archive)" },
    "investigator-api":      { "l1_hot": 8510,  "l2_verbatim": 0, "l2_index": 0, "l2_gist": 0, "rollout": "n/a (empty archive)" },
    "database-agent":        { "l1_hot": 5917,  "l2_verbatim": 0, "l2_index": 0, "l2_gist": 0, "rollout": "n/a (no archive)" }
  }
}