OpenClaw host briefing — AINL v1.2.8 / v1.3.0 / v1.3.1 / v1.3.2 / v1.3.3 (install / update)

Audience: You are the OpenClaw host — the process that owns config, cron, workspace paths, and what gets injected into the agent at session start.

Companion docs: OPENCLAW_AINL_GOLD_STANDARD.md (schedules + checklist) · WORKSPACE_ISOLATION.md · TOKEN_CAPS_STAGING.md · AGENT_AINL_OPERATING_MODEL.md · ../INTELLIGENCE_PROGRAMS.md · ../OPENCLAW_INTEGRATION.md

---

What the AINL v1.2.8 / v1.3.0 / v1.3.1 / v1.3.2 / v1.3.3 repo includes (you consume; you don’t reimplement)

| Capability | What it is | |------------|------------| | Evidence + sizing | ainl bridge-sizing-probe (and scripts/bridge_sizing_probe.py) samples SQLite namespace counts and sizes of ## Token Usage Report sections in daily memory, and suggests an AINL_BRIDGE_REPORT_MAX_CHARS target. CI exercises the probe. | | Rolling budget → intelligence cache | After the bridge publishes workflow / budget.aggregate / weekly_remaining_v1 (SQLite memory_records row; same record id as the legacy table name), scripts/run_intelligence.py merges that into MONITOR_CACHE_JSON under workflow → token_budget before each run (unless AINL_INTELLIGENCE_SKIP_ROLLING_HYDRATE=1). JSON output includes budget_hydrate. This matches a single aggregate read instead of scanning many days of markdown. ainl install openclaw still creates the legacy weekly_remaining_v1 table for compatibility; operational truth for rolling JSON is the memory row. | | Caps documentation | TOKEN_CAPS_STAGING.md and TOKEN_AND_USAGE_OBSERVABILITY.md: staging order — bridge (AINL_BRIDGE_REPORT_MAX_CHARS), then gateway (PROMOTER_LLM_*), with measurement first. | | Named env profiles | tooling/ainl_profiles.json + ainl profile list | show | emit-shell — dev, staging, openclaw-default, cost-tight so installs share a baseline without one-off env drift. | | Operator docs | Embedding pilot, WASM notes, TTL tuner, workspace isolation, HOST_PACK_OPENCLAW.md, and AGENT_AINL_OPERATING_MODEL.md (agent vs AINL roles; curated bootstrap must be loaded). | | Intelligence programs | In-tree: intelligence/token_aware_startup_context.lang writes .openclaw/bootstrap/session_context.md; intelligence/proactive_session_summarizer.lang summarizes prior days. Run via python3 scripts/run_intelligence.py context | summarizer | … |

Upgrade path: pip install -U 'ainativelang[mcp]' (or editable install from this repo), then ainl install-mcp --host openclaw if you use MCP. See ../INSTALL.md for RUNTIME_VERSION / __pycache__ after upgrades.

---

What the OpenClaw host must do (explicit)

1. Host contract — non-negotiable for chat-layer savings

When startup context has run, prefer injecting session_context.md (or the same curated path your layout uses) into the agent session instead of always loading the full MEMORY.md for bootstrap.

If the host ignores the curated file, token savings from AINL never appear in the chat layer — you only update files and subprocesses the model never sees.

See AGENT_AINL_OPERATING_MODEL.md. Prefer an upstream-supported bootstrap order; patching node_modules is fragile across OpenClaw upgrades.

2. Scheduling

• Cron / jobs that actually run:
  • python3 scripts/run_intelligence.py context (and summarizer on your chosen cadence — see gold standard for suggested times).
  • Bridge: at minimum weekly token trends (or equivalent) so rolling_budget_publish can write the rolling aggregate to memory_records (workflow / budget.aggregate / weekly_remaining_v1). (Legacy weekly_remaining_v1 table remains optional/secondary; install bootstraps it empty.)
• Use the same OPENCLAW_WORKSPACE, OPENCLAW_MEMORY_DIR, AINL_MEMORY_DB, MONITOR_CACHE_JSON (and siblings from WORKSPACE_ISOLATION.md) as run_wrapper_ainl.py and intelligence — one workspace, one truth.

3. Environment

• Operators can eval "$(ainl profile emit-shell openclaw-default)" (or cost-tight after measuring) for baseline AINL_* flags.
• Set AINL_WEEKLY_TOKEN_BUDGET_CAP if you want weekly_remaining_tokens in rolling JSON to match your real budget.
• Optional (for extra startup savings): enable embedding top-k candidate selection by ensuring:
  • AINL_STARTUP_USE_EMBEDDINGS=1 (profiles set this by default), and
  • AINL_EMBEDDING_MODE != stub (set to openai on the embedding/indexing cron process), and
  • you run the embedding pilot once (openclaw/bridge/run_wrapper_ainl.py embedding-memory-pilot) so embedding_workflow_index has refs, and
  • proactive_session_summarizer has run at least once so workflow.session_summary payloads include the real terse bullet text in payload.summary.
• Gateway-only: set PROMOTER_LLM_MAX_PROMPT_CHARS / PROMOTER_LLM_MAX_COMPLETION_TOKENS on the Apollo / gateway process per TOKEN_CAPS_STAGING.md (staging order — measure first).

4. Verification

• After a weekly bridge run: confirm the aggregate exists — either inspect memory_records for workflow / budget.aggregate / weekly_remaining_v1, and/or run ainl status (legacy table when non-null, else memory_records fallback for Weekly budget remaining). Confirm run_intelligence.py prints budget_hydrate with ok: true (not permanently skipped / no_rolling_record only). You may still verify the legacy weekly_remaining_v1 table with sqlite3 (legacy table; modern data lives in memory_records).
• Use ainl bridge-sizing-probe --json before tightening AINL_BRIDGE_REPORT_MAX_CHARS.

---

Bottom line

The AINL v1.3.3 tree delivers measurement, rolling-budget → cache hydration, profiles, caps staging docs, and clear host responsibilities (built on v1.3.0 OpenClaw/Hermes polish and v1.2.8 intelligence ops).

The remaining gap for ~85–90% usage/cost savings on session bootstrap is host behavior: load curated context + run scheduled intelligence + bridge jobs on shared paths. Without that, savings stay in files and subprocesses the model never reads.

Optional next tier (pilots, one at a time): EMBEDDING_RETRIEVAL_PILOT.md, WASM_OPERATOR_NOTES.md, TTL_MEMORY_TUNER.md.

---

Agent discovery

• tooling/bot_bootstrap.json → openclaw_ainl_gold_standard (checklist) · openclaw_host_ainl_1_2_8 (this briefing)