Subconscious

Autonomous, bounded bias layer that continuously evolves Alfred’s operating behavior by reinforcing, promoting, and governing self-improvement learnings and...

Security Scan

The skill's code and installer will install persistent cron jobs that autonomously ingest and promote 'learnings' into agent biases, which contradicts its own 'do not invoke proactively' guidance and warrants caution before installing.

Purpose & Capability

The skill claims to be a bounded, non-proactive bias layer but the package includes an installer that sets up recurring cron jobs (*/5 tick, hourly rotate with --enable-promotion, daily review, weekly benchmark). That installer therefore enables autonomous, persistent operation and auto-promotion of learnings — behavior not aligned with the SKILL.md top-line guidance to not invoke proactively. The package also bundles a second skill (self-improving-agent) inside the repo; bundling is plausible for convenience but increases the footprint and provenance concerns given the registry metadata lists no install spec and source is 'unknown'.

Instruction Scope

The runtime instructions and code read and write files under the OpenClaw workspace (e.g., ~/.openclaw/workspace-*/.learnings/ and memory/subconscious/). That is expected for a memory/bias system, but there is a clear contradiction: SKILL.md warns 'Do NOT invoke proactively' while other docs and the installer explicitly recommend and install automatic tick/rotate cron jobs and even call the learnings bridge automatically. The skill scans project learnings files and can promote items to live biases which are then injected into session prompts — this is powerful and should only run with explicit user consent, but the installer opts for automatic scheduling by default.

Install Mechanism

Registry metadata shows 'no install spec (instruction-only)', but the package contains a full installer script (scripts/install.sh) that modifies the user's crontab, creates directories/files under the detected workspace, and writes log files. The installer attempts to auto-detect the workspace and non-interactively edits crontab entries, enabling auto-promotion. Installing will therefore create persistent scheduled tasks on the host — a higher-risk install action than an instruction-only skill implies.

Credentials

The skill declares no required environment variables or credentials (only an optional SUBCONSCIOUS_WORKSPACE override). However, it reads and writes workspace files (.learnings/*, memory/subconscious/*) which may contain user-provided content. The self-improving-agent component explicitly warns not to log secrets, but the skill will index whatever is present in .learnings; if those files contain sensitive data, they could be read and turned into biases. No remote endpoints or secret-env-vars are requested, which is proportionate, but the file-level access to workspace content is noteworthy.

Persistence & Privilege

Although the skill is not marked 'always: true', the installer grants it persistent presence by adding cron jobs that run every 5 minutes and hourly (with --enable-promotion) and weekly benchmarks. Those cron jobs enable autonomous modification of the memory stores and promotion of biases without further manual invocation. This persistent autonomous behavior contradicts the 'do NOT invoke proactively' advisory in SKILL.md and increases the blast radius if a promoted bias is undesirable.

What to consider before installing

Before installing, consider the following: - Inspect scripts/install.sh line-by-line and do not run it blindly. That script will edit your crontab and create scheduled jobs that run every 5 minutes and hourly (the hourly job uses --enable-promotion by default). - If you want manual control, remove the --enable-promotion flag from the rotate cron entry (or don't install cron entries at all) so promotions only happen when you explicitly run the rotate command. - Backup your current crontab and OpenClaw workspace before running the installer (crontab -l > crontab.before) so you can restore it if needed. - Audit your workspace .learnings/ files for any sensitive information before connecting this skill; the skill will read those files and can convert learnings into injected biases. - Confirm you trust the package source and bundled self-improving-agent; registry metadata lists source as unknown and homepage absent — provenance is weak. Consider obtaining the package from a vetted source or inspecting all included code (the repo contains many Python modules that manage persistence and promotion logic). - If you prefer less autonomy, run the metabolism scripts manually (tick/rotate/review) from the scripts/ directory rather than installing cron jobs. - Verify hooks are not enabled: the package includes optional hooks for the self-improving-agent but they are disabled by default; do not copy them into your OpenClaw hooks directory unless you understand the implications. Given the contradiction between 'do not invoke proactively' in the SKILL.md and the installer's automatic cron + auto-promotion behavior, treat this package with caution and prefer a manual installation path or remove the scheduled auto-promotion before enabling it.

Subconscious v1.5 — Bounded Self-Improving Agent

What It Is

A persistent, self-evolving bias layer that survives session resets. Alfred's "second brain" — quiet, bounded, and strictly governed. It does NOT make decisions. It shapes how Alfred approaches decisions.

Core principle: Every mutation is typed, bounded, and logged. Core identity is untouchable without manual override.

Architecture

Learnings Bridge          Pending Queue           Live Store           Session Context
.learnings/ ─────────────► tick ──────────────► rotate ──────────────► bias inject
(Self-improving agent)    (reinforce,           (promote eligible,       (5 items max,
                         dedupe,               archive stale)           ephemeral)
                         skip duplicates)

Core Store ────────────────────────────────────────────────────────────────► Identity
(Immutable values,                                                  (never changes
 values/style, guiding principles)                                   without human OK)

Three layers:

• core/ — Immutable identity (Alfred's nature, Edward's preferences). Manual-only.
• live/ — Active learnings from experience. Governed promotion from pending.
• pending/ — Queue for new items. Bounded reinforcement before promotion eligibility.

Five bias types:

Kind	Category	Example
VALUE	Identity	"Alfred is sharp, calm, direct"
LESSON	Context	"XHS MCP needs QR re-login each cycle"
PRIORITY	Active	"Verify before claiming success"
PATTERN	Interpretation	"Proof discipline failure mode"
CONSTRAINT	Attention	"Don't suggest without trying first"

Lifecycle Commands

All commands run from scripts/ directory inside the skill:

cd ~/.openclaw/skills/subconscious/scripts

# Check system health
python3 subconscious_metabolism.py status
python3 subconscious_cli.py verify

# See active biases in session context
python3 subconscious_cli.py bias

# Manual metabolism cycles
python3 subconscious_metabolism.py tick        # Light tick (5 min cadence)
python3 subconscious_metabolism.py rotate       # Full rotation (hourly)
python3 subconscious_metabolism.py review       # Daily health check

Metabolism Cycles

Tick (every 5 min)

python3 subconscious_metabolism.py tick

• Scans .learnings/ via learnings bridge → queues new items to pending
• Passive reinforcement: increments reinforcement count on pending items
• Skips already-reinforced items (prevents type corruption)
• Runs bounded maintenance: freshness decay, metrics, no structural changes

Rotate (hourly)

python3 subconscious_metabolism.py rotate --enable-promotion

• Full maintenance: flush, decay, snapshot rotation
• Promotion gate (only with --enable-promotion):
  • confidence >= 0.75
  • reinforcement >= 3
  • freshness >= 0.3
  • Not a duplicate of anything in core/live/pending
  • Passes governance check
• Governance enforces typed mutation bounds on all items

Review (daily at 6am)

python3 subconscious_metabolism.py review

• System health check
• Snapshot integrity
• Pending queue depth check
• Recommendations log

Learnings Bridge

The learnings bridge connects the self-improving-agent skill to the subconscious:

• Scans .learnings/LEARNINGS.md, .learnings/ERRORS.md, .learnings/FEATURES.md
• Tracks seen entries per file in learnings_bridge_last_seen.json
• New entries → queued to pending.jsonl with type candidate_queued
• Entries already reinforced in this session → skipped (idempotent)

Bridge to self-improving agent: The learnings bridge should be called from the metabolism tick so that new learnings flow into the subconscious every 5 minutes automatically.

Checking System Health

python3 subconscious_metabolism.py status

Expected output:

Core: 3/50   Live: 1/100   Pending: 0/500   Snapshots: 10/10
Status: OK

If Pending: 0/500 and Live: N — system is healthy, items promoting correctly. If Status: blocked — resource limits hit, run rotate to compact.

Adding Items Manually

# Queue a lesson manually
python3 subconscious_cli.py intake --kind LESSON --text "Remember to verify before claiming success" --confidence 0.8 --source "manual"

# Check what's in pending
python3 subconscious_cli.py pending

Key Files

File	Purpose
subconscious/schema.py	Item dataclasses, validation, kind enum
subconscious/store.py	JSON file ops with atomic writes
subconscious/retrieve.py	Relevance scoring, is_duplicate
subconscious/influence.py	Convert items to bias blocks for prompts
subconscious/governance.py	Mutation types, protection classes, bounds
subconscious/evolution.py	Promotion pipeline, reinforcement logic
subconscious/maintenance.py	Decay, snapshot rotation, housekeeping
subconscious/intake.py	Conservative item extraction from turns
subconscious/flush.py	Snapshot building/loading for session continuity
subconscious/learnings_bridge.py	Bridge to self-improving-agent .learnings/

For Claude Code Sessions

When Claude Code needs to assess or improve the subconscious system, use the claude-cx wrapper:

claude-cx "Run: python3 ~/.openclaw/skills/subconscious/scripts/subconscious_metabolism.py status"
claude-cx "Run: python3 ~/.openclaw/skills/subconscious/scripts/subconscious_cli.py bias"
claude-cx "Read ~/.openclaw/skills/subconscious/subconscious/evolution.py lines 450-525"

Cron Setup

After running install.sh, four cron jobs are active:

• */5 * * * * — tick (light metabolism)
• 0 * * * * — rotate (hourly, with promotion)
• 0 6 * * * — review (daily health check)
• 0 9 * * 1 — weekly benchmark (Monday 9am, compares to baseline)

Weekly Benchmark

Compares current state to the baseline snapshot captured at install time. Reports on:

• Learnings volume change (LEARNINGS.md, ERRORS.md entries)
• New biases promoted to live
• Pending queue depth
• Error recurrence as a proxy for learning effectiveness

# Run manually
python3 scripts/subconscious_benchmark.py

# View past benchmarks
ls memory/subconscious/benchmarks/
cat memory/subconscious/benchmarks/benchmark_YYYY-MM-DD.json

To capture a new baseline:

python3 scripts/subconscious_benchmark.py --capture-baseline

To override workspace location:

export SUBCONSCIOUS_WORKSPACE=/path/to/any/workspace