メニュー
Durable cross-session memory for claudemax. SQLite+FTS5 hybrid-retrieval store with 5-tier CoALA taxonomy, multi-scope identity, and staleness handling. Use `recall` before claiming prior-session facts, `add` to record durable lessons, `verify` to refresh a recalled row.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Ask. Achieve. The canonical user-facing entry to claudemax — describe your goal, the SOTA-2026 pipeline (deepresearch → multispec → parallel /goal → verify) handles the rest. Same engine as /cmax with a friendlier verb. CLI equivalent is `cmax ask "<goal>"`.
The default FAT umbrella. Auto-runs deepresearch + multispec decomposition + parallel /goal + verify. Opus for planning/judgment/verification; Sonnet for sub-Spec execution (escalates per router). Multispec engine selects Mode A (SDK subagents) or Mode B (Claude Code Agent Teams) automatically based on work size. Your default daily-driver.
Max-effectiveness umbrella. Opus for EVERYTHING — planning, decomposition, sub-Spec /goal execution, verification. Use for novel domains, security/auth/payments, or when you want the highest ceiling regardless of cost.
Show the model-routing decision for a task (Fable/Opus/Sonnet/Haiku) with reasoning. Use before launching a packet to sanity-check the tier and to override when you have stronger judgment than the heuristic.
Reviews a proposed diff for correctness, security, style, taste alignment. Distinct from /audit (which reviews existing code, not diffs) and /verify (which checks SPEC completion).
Spec quality gate. Every completion condition must have a mechanically-checkable verifyHint (file path, command, test name, behavior). Blocks /goal handoff if the spec wouldn't be verifiable.
| name | memory |
| description | Durable cross-session memory for claudemax. SQLite+FTS5 hybrid-retrieval store with 5-tier CoALA taxonomy, multi-scope identity, and staleness handling. Use `recall` before claiming prior-session facts, `add` to record durable lessons, `verify` to refresh a recalled row. |
Hybrid-retrieval SQLite+FTS5 store built on the SOTA-2026 evidence base. Replaces v1's flat-file Obsidian mirror; anti-pattern to dual-write into a Markdown tree (Letta 2025-08-12 finding: filesystem alone is sufficient).
cmax memory recall "<task>" [--depth simple|medium|deep] [--lane LANE] [--run RUN] [--no-stale]
cmax memory add <tier> "<content>" [--title T] [--tags T1,T2] [--lane LANE] [--run RUN]
cmax memory verify <source>#<id> [--by <agent-name>]
Plus utility commands search, runs, credit, drain (queue → SQLite).
| Situation | Verb |
|---|---|
| "Did we decide X already?" / "What did we learn about Y last week?" | recall first. Quote the row verbatim or say "no recall". |
| You made a non-obvious architectural call. | add decision "<body>" --lane <id> |
| You debugged an error whose fix wasn't obvious from the stack trace. | add error-solution "<sig> :: <fix>" |
| You discovered a code pattern that's worth reusing. | add pattern "<name>" "<body>" |
| A general note about this session's work. | add episode "<one-line note>" |
| You followed a recalled row and it turned out correct. | verify <source>#<id> --by <agent> |
| Lane-spawn moment writes to memory; want async to avoid blocking. | append to .claudemax/memory.queue.jsonl, then cmax memory drain later |
| Depth | Limit | Use |
|---|---|---|
simple | 5 hits | Quick "is there anything on X?" check |
medium | 15 hits (default) | Standard session-start recall |
deep | 50 hits | Investigations / archaeology |
Composite score is BM25 + entity-match boost + recency decay + recently-verified boost − staleness penalty. Mem0's 2026-05-21 SOTA report identifies multi-signal hybrid retrieval as a winning architecture vs pure semantic similarity.
Based on arxiv 2309.02427 (Cognitive Architectures for Language Agents). The harness maps each tier to a SQLite table:
| Tier | Table | What goes in |
|---|---|---|
episode | episodes | Session-scoped events: "started X", "blocked on Y", "session ended with Z passing" |
decision | decisions | Architectural calls + their rationale |
pattern | patterns | Recurring named code patterns + the body |
error-solution | errors_solutions | Stack-trace-like signature + the fix |
graph | episodes (kind='graph') | Multi-step stories with relationships |
Every memory row carries optional run_id, lane_id, user_id, app_id. This is the Mem0 winning architecture (May 21 2026). Use these to:
cmax memory recall "auth" --lane infra-lane-implement-apps-webhook-dispat--run run-1779392855991Rows whose last_verified_at is older than 30 days (configurable) get a [STALE] flag on recall. Calling cmax memory verify <ref> refreshes the timestamp and bumps a verified-count. This is the harness's answer to the #1 unsolved problem in 2026 agent memory (Mem0 SOTA report).
--no-stale excludes stale rows entirely. Default is to include them with the flag so you can decide.
Subagent system prompts in v0.2.3+ include a memory tool rule:
Before claiming any fact about prior sessions, prior decisions, or "what we did last time", call
cmax memory recall "<topic>" --depth mediumfirst.
Backed by Letta's 2025-08-12 finding that agents constrained by tool rules outperform unconstrained agents (LoCoMo 74.0% vs 68.5%). Complements the dark-patterns no-fake-recall.sh hook which blocks "as we discussed earlier"-shaped claims at Stop time.
Separate from durable memory but adjacent. When you spawn cmax mega --features-checklist, each lane gets a .claudemax/lanes/<lane-id>/features.json checklist. Subagents:
passes: false.passes: true. Exit.JSON not Markdown because per Anthropic: "the model is less likely to inappropriately change or overwrite JSON files."
git log / git blame..no-amnesia/state/CURRENT.md working-state for that).All accessed 2026-05-21: