| name | memory-recall |
| description | Auto-activate on any question about past work, decisions, commands, facts, or current status ("what did we do this morning/yesterday/today", "how\ |
| s everything looking", "what\'s in flight", "remember when", "have we seen this error before", "what is the IP of X"). Encodes R… | status/workboard protocol, time-bounded browse first, multi-angle search + trigram fallback, cross-agent awareness.' |
| tags | ["rivetos","memory","recall","discipline","rivet-memory","status"] |
| version | 0.3.0 |
RivetOS Memory Recall Discipline (Grok Build)
Shared persistent memory across every Rivet agent (rivet-claude, rivet-hermes, grok, etc.) via the RivetOS MCP server.
Usual failure mode is poor discipline — wrong tool, wrong evidence layer, or a browse flooded by tool rows.
The Core Rules
0. Status / "how's things" question? → Workboard first (not host health)
Prompts like "how's everything looking", "what's going on", "status", "what's in flight", "catch me up" mean current work across agents, not memory_stats or box uptime.
memory_browse(window="last_24h") (default already excludes tool rows) — read user + assistant closers from all agents.- If thin, widen (
today / yesterday) or memory_search for the live thread (PR numbers, "merged", "blocked", "hands off").
- Treat
workspace/memory/*.md and HEARTBEAT dated notes as hints to verify, never as current state. RivetOS conversation memory (and session-state / wiki when present) wins.
- Host/
memory_stats only as a secondary note when relevant — never the lead answer.
1. Time-bounded question? → memory_browse with window= FIRST
Any prompt that pins a timeframe ("this morning", "yesterday", "today", "earlier", "last week", "the standup") needs chronological results, not relevance rank.
Use window= when available:
window="this_morning" | "yesterday" | "today" | "this_week" | "last_24h" | "last_7d" | "last_14d"
Prefer rolling last_7d / last_14d over this_week early in the week. Fallback: full UTC ISO since / before (never bare dates).
After a browse hits the limit, raise limit (max 200) or flip order — do not assume you have everything.
Default browse excludes role=tool. Limit budget goes to user/assistant/system. Pass include_tools=true only when debugging capture, harness wiring, or tool failures.
2. Topic / lookup question (no clear timeframe)? → Multi-angle search, minimum 3 queries
Vary by service/role, host/nickname, network, and exact tokens. Use mode="trigram" for IPs, MACs, error strings, hostnames. FTS OR/phrases when supported.
3. Semantic/FTS returns thin? Immediately retry with mode="trigram"
4. "No results" / user pushback → change strategy, do not give up
Widen window, add angles/trigram, check other agents. Only then go external.
Decision Flow
User asks about past / status / "remember" / facts
│
▼
Status / how's things / in flight?
YES → workboard browse (rule 0)
NO
│
▼
Mentions clear timeframe?
YES NO
│ │
▼ ▼
memory_browse 3× memory_search
(window= preferred) (different angles)
│ │
▼ ▼
Still thin? Thin results?
Widen / raise limit Retry trigram + FTS
Grok Build Tool Usage
search_tool to discover qualified names (rivetos__memory_browse, etc.).- Time-bounded →
window= on browse.
- Status → rule 0 before
memory_stats.
Cross-Agent Reality
Hits from rivet-claude / rivet-hermes / grok are equally valid. Filter by agent only when the user wants one lineage.
Why This Exists
2026-05-23 WAP-DHCP: keyword search returned nothing; memory_browse over the morning window recovered the full incident. 2026-08-25 status miss: memory_stats + stale memory/*.md buried a live Claude workboard (#545 merged, on-device validation).
Related
memory-today / memory-yesterday — thin wrappers around rule 1 with the matching window.memory_stats — coverage/health diagnostics, not a status briefing.- Pre-compaction Hermes captures are high-value for long sessions.
