Jun 22, 2026 · 10:47 AM CDTHealthy but noisyHermes memory stack

Hermes memory foundation audit

Recent sessions show a working memory stack with unclear ownership boundaries. The fix is not replacement. It is sharper contracts, targeted cleanup, and shorter skills in volatile domains.

Owner read

Keep LCM as raw/exact context, keep flat memory as the tiny always-injected behavior card, keep Mnemosyne as durable extracted recall, and keep skills as procedures. The current risk is cross-layer residue: Mnemosyne has old transcript noise, flat memory has facts Mnemosyne cannot recall, and skills contain stale implementation snapshots.

100%
LCM FTS sync
65/65
Mnemosyne episodic vectors
1,595
historical conversation rows in Mnemosyne
0
recent session_search calls found

Live wiring

Context enginelcm
Memory providermnemosyne
Built-in memoryalways active
Compression modelgemini-3-flash-preview
LCM plugin0.16.1

DB health

Mnemosyne quick_checkok
LCM quick_checkok
LCM messages / FTS25,926 / 25,926
LCM summaries / FTS68 / 68
Expired Mnemosyne rows0

Clear gaps

Fix now

  • Mnemosyne contains prompt-echo episodic artifacts such as “Please provide the memories…” with high recall counts.
  • Flat memory contains durable facts with zero Mnemosyne hits, including ADHD, config-driven, Spectrum-ts v5.2.0, and cache_friendly.
  • session_search is underused for “what happened before?” despite Hermes docs positioning it as the right historical-recall path.

Watch

  • Mnemosyne raw-turn sync is intentionally disabled by catch-all ignore. Good for bloat, bad for passive learning.
  • LCM summaries can contain stale operational facts. Treat them as session compression, not canonical truth.
  • USER.md and MEMORY.md are near capacity and carry sensitive always-injected details.

Layer contract

Correct ownership

Flat memorytiny behavior card
Mnemosynedurable extracted facts
LCM/session DBraw history + exact recovery
Skillsprocedures + runbooks
Brain/Obsidianexplicit human-facing notes

Recall rule

  • Use session_search first for cross-session “where did we leave off?” questions.
  • Use LCM tools for current-session compression and exact expansion after a session is identified.
  • Use Mnemosyne for stable preferences/facts that should influence future behavior.

Skill audit

Clean up

  • mnemosyne-operations: highest-priority prune. Keep current principles; move version-specific forensics to references.
  • hermes-local-operations: remove hardcoded LCM version/config snapshots; replace with verification commands.
  • hermes-agent: too broad; make it docs-first and move local recipes elsewhere.
  • native-mcp: fix hot-reload/current CLI guidance.

Keep as model

  • here.now: right pattern for changing domains — concise local workflow, current docs, live API truth.
  • kanban-worker: appropriately scoped to durable pitfalls; lifecycle lives in injected guidance.

Recommended actions

1. Prune reviewed Mnemosyne garbage.Back up first. Delete prompt echoes and fabricated examples. Clean vector rowids after episodic deletes.
2. Promote durable flat-memory facts into Mnemosyne intentionally.Use explicit source and veracity. Keep flat memory only for always-needed behavior.
3. Add a recall discipline rule.Before repeating Hermes/debug work, call session_search; then use LCM for exact expansion when needed.
4. Prune volatile skills.Short top-level skill, links to official docs, local verification commands, detailed forensics in references.
5. Reduce sensitive always-injected profile details.Keep only details that actively improve behavior every session.

Evidence

Counts

Working memories1,670
Unconsolidated backlog21
Unknown veracity working rows1,612
Never recalled working rows918
Prompt-echo episodic artifactspresent

Recent-session behavior

  • Mnemosyne recall was mostly used in explicit memory-audit sessions.
  • LCM was used effectively for compaction checks and exact recovery.
  • Agents sometimes guessed or asked before loading the right skill or searching session history.