heleni-memory-architecture（海伦尼内存架构）

Load Local Context

CONTEXT_FILE="/opt/ocana/openclaw/workspace/skills/memory-architecture/.context"
[ -f "$CONTEXT_FILE" ] && source "$CONTEXT_FILE"
# Then use: $WORKSPACE, $MEMORY_FILE, $WHATSAPP_MEMORY_DIR, $OWNER_TIMEZONE, etc.

Memory Architecture Skill (Honcho-inspired)

Core Principle

Two types of memory, three tiers. Inspired by Honcho/Hermes Agent memory system.

Memory Types

[FACT] — Explicitly stated

Something the owner said directly.

"Netanel works until 20:00" → [FACT]

[DEDUCED] — Logical conclusion from behavior

Something inferred from patterns, corrections, or repeated behavior.

"Netanel prefers execution over explanation" → [DEDUCED] (inferred from repeated feedback pattern)

Rule: After any interaction where a pattern becomes clear — write a [DEDUCED] entry. Don't wait for the owner to say it explicitly.

Memory Tiers

Project Files

For active, multi-session work (rollouts, ongoing tasks, recurring topics), maintain a project file:

memory/projects/pa-rollout.md
memory/projects/crons-health.md
memory/projects/my-project.md

Each project file should contain:

• Status — current state
• Open loops — unresolved items that carry across days
• Key decisions — what was decided and why

Load on demand: Read the relevant project file when that topic comes up — not every session. Daily files are raw log. Project files are living context.

Rule: If a task or decision needs to be remembered tomorrow — put it in the project file, not just the daily log.

Session Summary Rule

At the end of any conversation with 5+ meaningful exchanges, append to memory/YYYY-MM-DD.md:

## Session Summary — HH:MM

**Decisions made:**
- [list]

**Tasks completed:**
- [list]

**Deduced patterns:**
- [DEDUCED] [observation]

**Promote to MEMORY.md?**
- [ ] Yes: [what]
- [x] No

When to promote to MEMORY.md:

• Pattern has appeared 2+ times across different sessions
• Owner explicitly corrected something → update MEMORY.md immediately
• Important preference discovered → add as [DEDUCED]

Writing Deduced Memories

After a correction or repeated behavior, write:

- [DEDUCED] <conclusion> — evidence: <what happened>

Examples:

• [DEDUCED] Prefers English for work docs — corrected me when I created Hebrew doc (2026-04-03)
• [DEDUCED] Expects autonomous execution without asking permission for reversible tasks — never asks why I acted autonomously
• [DEDUCED] Reads messages outside work hours but won't always respond

What NOT to remember

• One-time events with no pattern value
• Transient task state (use monday.com for that)
• Secrets or credentials (never in memory files)
• Conversation noise

Memory Quality Checklist

Before writing to MEMORY.md, ask:

• [ ] Is this a fact OR a deduced pattern (not just noise)?
• [ ] Will this be useful in a future session?
• [ ] Is this tagged [FACT] or [DEDUCED]?
• [ ] Does it contradict an existing entry? (If yes, update the old one)

Self-Review Cron Loop

For continuous self-improvement, set up two daily crons:

Midday (13:00 local):

openclaw cron add \
  --name "heleni-midday-self-review" \
  --cron "0 13 * * *" \
  --tz "Asia/Jerusalem" \
  --session isolated \
  --message "Run midday self-review: read today's memory file, find mistakes or corrections from this morning, update MEMORY.md with [DEDUCED] lessons if needed, commit to git. Silent if nothing to fix." \
  --timeout-seconds 180

Nightly (23:00 local):

openclaw cron add \
  --name "heleni-internal-self-review" \
  --cron "0 23 * * *" \
  --tz "Asia/Jerusalem" \
  --session isolated \
  --message "Run nightly self-review: read today's full memory file + last 3 MEMORY.md entries. Find: repeated mistakes, patterns to promote to long-term memory, rules to tighten. Update MEMORY.md and/or relevant SKILL.md files. Commit to git with message 'Self-review YYYY-MM-DD: [what changed]'. Silent if nothing to improve." \
  --timeout-seconds 180

Why twice a day:

• Midday catches morning mistakes early — fixes them before they repeat in the afternoon
• Nightly does the full consolidation of the day
• Once a day (evening only) is too slow for fast-moving days

Memory Compaction (Prevent Bloat)

Left unchecked, MEMORY.md and AGENTS.md grow unbounded — causing context window bloat and degraded performance.

Targets:

• MEMORY.md → max 175 lines
• AGENTS.md → max 60 lines

When to compact:

• Weekly — set up a cron (see below)
• When MEMORY.md exceeds 200 lines
• When loading files already consumes >5% of context window

How to compact:

1. Read MEMORY.md fully
2. Merge duplicates, remove outdated entries, trim examples
3. Keep: active rules, key contacts, deduced patterns still valid
4. Remove: removed crons, resolved issues, one-time events
5. Commit to git after

Weekly compaction cron:

openclaw cron add \
  --name "weekly-memory-compaction" \
  --cron "0 7 * * 0" \
  --session isolated \
  --model "anthropic/claude-haiku-4-5" \
  --message "Weekly memory compaction: Read MEMORY.md and AGENTS.md. Remove outdated entries, merge duplicates. Target: MEMORY.md <175 lines, AGENTS.md <60 lines. Git push after. NO_REPLY." \
  --announce \
  --timeout-seconds 120

Signal you need compaction:

• MEMORY.md > 200 lines
• Session startup feels slow
• You notice you're repeating old rules that are no longer relevant

Production Notes

• Workspace path: /opt/ocana/openclaw/workspace (not ~/.openclaw/workspace)
• WhatsApp memory path: memory/whatsapp/groups/<JID-sanitized>/context.md or memory/whatsapp/dms/<PHONE-sanitized>/context.md
• Sanitize rule: replace @, ., + with -
• After writing memory — push to git (git-backup skill)
• "sure thing" is a behavior rule (not a memory item) — it lives in SOUL.md, do not re-derive from memory

Cost Tips

• Cheap: File writes are free — don't hesitate to log
• Session summaries: Write once at the end, not after each message
• MEMORY.md promotions: Batch once per session, not per interaction