Error Memory — learning from errors

Trigger

1. At session start — load ERROR_LOG.md, remember all errors

2. After ANY error — write to ERROR_LOG.md

3. Before action — verify this error was not already encountered

File

~/workspace\.openclaw\workspace\_memory\ERROR_LOG.md

Important: The file is created automatically on the first error. If it doesn't exist — create it via mkdir -p ~/workspace/.openclaw/workspace/_memory and write a header.

Read on startup. Update after errors.

Session start check:

ls ~/workspace/.openclaw/workspace/_memory/ERROR_LOG.md 2>/dev/null || echo "ERROR_LOG.md not yet created"

Write format

## YYYY-MM-DD — Category: Brief description



**Problem:** What happened

**Cause:** Why (root cause)

**Solution:** How fixed / how to avoid

**Tool:** Which tool NOT to use / WHICH to use instead

**Signal:** How to recognize this error in future (keywords)



---

When to log

Categories:

• browser — captcha, block, Cloudflare, 404, datacenter IP

• tool — tool broke, not installed, wrong path

• api — API key, 403, rate limit, timeout

• command — wrong command, syntax, flag

• proxy — proxy, VPN, IP, DNS

• telegram — Telegram API, bot, sending, formatting

• file — path, encoding, permissions, size

• cron — cron job, schedule, delivery

• proxy-server — localhost:4000, OmniRoute, ConnectionResetError

• apk — APK, Android, decompilation, ADB

• install — pip, npm, uv, dependency

• env — environment: python vs python3, PATH, Windows quirks, jqAvailability

• search — search-specific: captcha_detected, engine failures, extract

Mandatory error checklist

1. Error occurred

2. → Why? (root cause)

3. → How to fix? (solution)

4. → How to prevent? (prevention)

5. → Log to ERROR_LOG.md

6. → Update memory if critical

7. → If error relates to a new tool → update TOOLS_INDEX.md

8. → Dedup: is there already an entry for this gotcha? If yes — update it, don't duplicate

Example errors to log

Browser via Browserbase = captcha

Problem: Cloudflare/DDG/API block requests

Cause: IP <IP> — datacenter A1 Minsk, all search engines ban

Solution: local browser + computer-use (local) OR search OR Happ VPN

Signal: Cloudflare, CAPTCHA, "blocked", 403 from search, "Attention Required"

Tavily 403

Problem: web_search via Tavily returns 403

Cause: datacenter IP, Tavily blocks

Solution: search (self-hosted) OR local browser OR HN Algolia API

Signal: 403, Tavily, "Web tools are not configured"

ConnectionResetError on proxy

Problem: proxy.py crashes with ConnectionResetError(10054)

Cause: Remote host resets connection (OmniRoute/upstream down)

Solution: Restart proxy.py OR check OmniRoute on localhost:4000

Signal: ConnectionResetError, 10054, proxy.py

On session start

# 1. Load ERROR_LOG

error_log = read_file("_memory/ERROR_LOG.md")



# 2. Read ALL errors — they're now in memory



# 3. Before each action:

#    "Has this problem happened before?" → search ERROR_LOG by keywords

Daily cleanup: deduplication and archival

Once per session (on startup or at end) — run ERROR_LOG.md

through cleanup, so the log stays a tool, not a dump:

1. Deduplication. One gotcha = one entry. Match by meaning/signal

   merge into one representative entry (class), originals → archive.

   Most common duplicates: same tool gotcha from different sessions

   (example: BACKGROUND not supported ×2, trash-multipath ×4).

2. Archive entries older than 3 days. Entries with date older than 3 days from current —

   verbatim to _archive/ERROR_LOG.<cleanup-date>.md (precedents:

   _archive/ERROR_LOG.2026-08-05.md, ERROR_LOG.2026-08-12.md). In the fresh

   file — link header: "> Archive for in _archive/ERROR_LOG..md.

   File recreated for fresh entries." Delete nothing — only move.

3. Exception: live gotchas. Age ≠ irrelevance. Entries about live

   infrastructure and reproducible environment gotchas (freebuff-importer, PORT, 401/429,

   BACKGROUND, Git Bash/PowerShell, trash, Preview, mcp.json,

   health-probes) keep in fresh log regardless of date.

4. Sanitization. Before commit/archive — mask personal data

   (email, IP, tokens): user***@gmail.com. Don't link side accounts

   to identity in git history.

5. Append technique. Append entries only via python by the exact

   last entry header (file is CRLF; str_replace anchors on tail

   break — see "append in ERROR_LOG" entry). During mass rebuild

   preserve verbatim text and line endings (newline="").

Anti-patterns

• ❌ Error occurred → continued working without logging

• ❌ Logged in thoughts — not in file (session dies, memory lost)

• ❌ Didn't read ERROR_LOG on startup — will repeat same errors

• ❌ Logged error without "Signal" — won't find it in search

• ❌ Logged "tool doesn't work" instead of "how to fix" — doesn't preserve the working fix