Hermes Session Maintenance

Use this skill when the user asks about pruning Hermes sessions, session retention, session DB size, hermes sessions prune, sessions.auto_prune, sessions.retention_days, session export/backup, or whether old conversations should be capped.

Operating principles

• Treat this as Hermes configuration/maintenance work. If broader Hermes setup or CLI behavior is involved, also load hermes-agent for authoritative command references.
• Do not assume retention policy from memory. Check live state before recommending destructive cleanup.
• Pruning is deletion. Prefer exporting a safety snapshot before the first aggressive prune or before reducing retention substantially.
• Distinguish default behavior from active behavior:
• hermes sessions prune defaults to --older-than 90.
• sessions.retention_days may also default to 90.
• sessions.auto_prune: false means nothing is being deleted automatically.
• Only ended sessions are prune candidates; active sessions are not pruned.

Standard workflow

1. Inspect current volume:

hermes sessions stats

Use total sessions, messages, per-source counts, and DB size to shape the recommendation.

2. Inspect prune command behavior when needed:

hermes sessions prune --help

3. Check current config values:

hermes config show

If the summary output does not include the session settings, inspect the config file's sessions: block directly. Do not use hermes config get; this CLI supports show, edit, set, path, env-path, check, and migrate, not get.

4. Recommend retention based on actual size and user needs.

• 90 days is conservative/default.
• 30 days is reasonable when the DB is multi-GB and session search is mostly used for recent operational continuity.
• Shorter than 30 days should usually be explicit user preference or a storage emergency.

5. Before first major prune, export a backup:

hermes sessions export ~/hermes-sessions-backup-$(date +%Y%m%d).jsonl

6. Enable auto-prune if approved:

hermes config set sessions.auto_prune true
hermes config set sessions.retention_days 30

7. Reclaim space immediately with a manual prune:

hermes sessions prune --older-than 30

Add --yes only if the user has authorized non-interactive deletion.

8. Verify after pruning:

hermes sessions stats

Source-specific cleanup

When one channel is the bloat source, prune only that source:

hermes sessions prune --older-than 30 --source cli
hermes sessions prune --older-than 90 --source telegram

Use this when the user wants to preserve messaging history longer than CLI scratch sessions.

For ongoing split retention, do not rely on global sessions.auto_prune because it has one retention window and can prematurely prune the longer-lived source. Use a quiet script under ~/.hermes/scripts/ and a local no-agent cron job instead:

hermes config set sessions.auto_prune false
hermes config set sessions.retention_days 45 # documentation/default only when auto_prune is off
# ~/.hermes/scripts/prune-sessions-by-source.sh:
# hermes sessions prune --older-than 15 --source cli --yes
# hermes sessions prune --older-than 45 --source telegram --yes

Schedule it daily with deliver='local' and no_agent=true so it does not message the user unless it fails. Cron script paths must be relative to ~/.hermes/scripts/, not absolute.

For the detailed split-retention pattern and one-time CLI purge procedure, see references/source-specific-retention-and-cli-clear.md.

One-time CLI clears

When the user explicitly asks to clear all CLI sessions, first try the CLI pruner:

hermes sessions prune --older-than 0 --source cli --yes

Then verify. If many CLI sessions remain, check whether they are unended/active (ended_at IS NULL). The built-in pruner only removes prune candidates, so historical CLI sessions that were never marked ended may remain.

For an explicit destructive CLI-only clear, keep the newest/current CLI session, delete messages and sessions rows for other source='cli' sessions directly in ~/.hermes/state.db, then checkpoint/VACUUM after stopping processes that hold the DB. Preserve Telegram/Discord/cron/subagent sessions unless separately requested. See the reference file above for the SQL and verification queries.

Pitfalls

• Do not present the 90-day default as a hard cap. It is a default cutoff for pruning/retention, not an immutable limit.
• Do not say auto-prune is active just because retention_days is set. sessions.auto_prune must be true.
• Do not skip hermes sessions stats; the right recommendation depends on live DB size and source mix.
• The live session store is typically ~/.hermes/state.db, not ~/.hermes/sessions.db. If reclaiming disk space manually, inspect state.db and its -wal/-shm companions.
• SQLite may not physically shrink after pruning while Hermes CLI/gateway/dashboard processes hold state.db open. A WAL checkpoint may need gateway/dashboard temporarily stopped; a full VACUUM may still require all Hermes CLI sessions to exit. Do not leave a large state.db-wal behind after attempting VACUUM — run PRAGMA wal_checkpoint(TRUNCATE); once locks are clear.
• Do not imply memory, skills, cron jobs, notes, or project files are affected by session pruning. Session pruning affects session transcripts/state only.
• Do not over-explain when the user asks for a go/no-go recommendation. Give the recommendation, the tradeoff, and the exact commands.
• For the user, session-retention questions usually want a direct recommendation first, not a long policy discussion. If live stats show a multi-GB DB and mostly recent operational use, recommend 30 days with auto-prune; mention 45 days as the conservative alternative and 90 days only for deliberate archive-style use.

Session forensics: token burn, suspected loops, and runaway turns

Use this skill not only for pruning/retention, but also when the user says a session "looped," burned too many tokens, felt unusually slow, or may have resent content repeatedly.

Fast diagnostic workflow

1. Read the session from the DB and identify whether the waste came from:

• repeated assistant/tool turns inside one user request,
• duplicate gateway delivery,
• oversized skill/tool payloads,
• or a genuinely long final answer.

2. Check session totals in state.db:

sqlite3 ~/.hermes/state.db <<'SQL'
.headers on
.mode column
SELECT id, source, title, model, input_tokens, output_tokens,
cache_read_tokens, reasoning_tokens, api_call_count, message_count,
datetime(started_at,'unixepoch','localtime') AS started,
datetime(ended_at,'unixepoch','localtime') AS ended
FROM sessions
WHERE id='SESSION_ID';
SQL

3. Inspect per-message shape to see the turn pattern and tool payload sizes:

sqlite3 ~/.hermes/state.db <<'SQL'
.headers on
.mode column
SELECT id, role, tool_name, tool_call_id, length(COALESCE(content,'')) AS content_len,
datetime(timestamp,'unixepoch','localtime') AS ts
FROM messages
WHERE session_id='SESSION_ID'
ORDER BY id;
SQL

4. Check gateway timing/delivery behavior:

• look for inbound message, response ready,
• and especially Suppressing normal final send ... content_delivered=True. That pattern usually means streaming already delivered the final answer once, so the problem was not duplicate send.

5. Attribute blame precisely:

• large skill_view payloads,
• multiple overlapping web_search / web_extract calls,
• repeated assistant-empty/tool turns,
• or unnecessary env dumps for simple lookup questions.

What to conclude

• If gateway logs show one inbound and one response ready, with content_delivered=True, call it an internal tool-call cascade or context-growth problem — not a transport resend loop.
• If a single loaded skill or search result dominates payload size, call that out explicitly. Quantify it.
• Distinguish input-token burn from output-token size. Often the answer is short; the cost came from replaying tool results back into the model.

Pitfalls

• Do not call every high-token session a "loop." Many are just over-eager research/tool fan-out.
• Do not blame Telegram delivery when gateway logs show suppression of duplicate final send.
• Do not stop at session_search() if the user is asking about token burn. Use state.db and gateway.log so the diagnosis is grounded.
• Do not save transient missing-binary issues as durable skill knowledge. The durable lesson is to avoid loading bulky skills or fanning out searches unnecessarily when a cheap-first path would answer the question.

Cheap-first guidance to recommend after diagnosis

For lightweight link-analysis or one-question sessions:

• prefer one direct extract/read path first,
• avoid loading a full skill unless the tool is truly needed for writes or auth-specific operations,
• cap redundant searches,
• and use minimal env queries for session-ID lookups instead of broad dumps.

See also: references/session-forensics-token-burn.md.

References

• references/session-pruning-retention.md — condensed notes from the session where 30-day auto-prune was evaluated against a 6.7GB session DB.
• references/session-forensics-token-burn.md — diagnosing whether a session burned tokens because of duplicate delivery, internal tool cascades, or oversized tool payloads.

Public support files

• references/session-forensics-token-burn.md
• references/session-pruning-retention.md
• references/source-specific-retention-and-cli-clear.md