Persistent Memory

Hermes Agent has bounded, curated memory that persists across sessions. This lets it remember your preferences, your projects, your environment, and things it has learned.

Two files make up the agent's memory:

Both are stored in ~/.hermes/memories/ and are injected into the system prompt as a frozen snapshot at session start. The agent manages its own memory via the memory tool — it can add, replace, or remove entries.

At the start of every session, memory entries are loaded from disk and rendered into the system prompt as a frozen block:

══════════════════════════════════════════════
MEMORY (your personal notes) [67% — 1,474/2,200 chars]
══════════════════════════════════════════════
User's project is a Rust web service at ~/code/myapi using Axum + SQLx
§
This machine runs Ubuntu 22.04, has Docker and Podman installed
§
User prefers concise responses, dislikes verbose explanations

The format includes:

• A header showing which store (MEMORY or USER PROFILE)
• Usage percentage and character counts so the agent knows capacity
• Individual entries separated by § (section sign) delimiters
• Entries can be multiline

Frozen snapshot pattern: The system prompt injection is captured once at session start and never changes mid-session. This is intentional — it preserves the LLM's prefix cache for performance. When the agent adds/removes memory entries during a session, the changes are persisted to disk immediately but won't appear in the system prompt until the next session starts. Tool responses always show the live state.

The agent uses the memory tool with these actions:

• add — Add a new memory entry
• replace — Replace an existing entry with updated content (uses substring matching via old_text)
• remove — Remove an entry that's no longer relevant (uses substring matching via old_text)

There is no read action — memory content is automatically injected into the system prompt at session start. The agent sees its memories as part of its conversation context.

Substring Matching

The replace and remove actions use short unique substring matching — you don't need the full entry text. The old_text parameter just needs to be a unique substring that identifies exactly one entry:

# If memory contains "User prefers dark mode in all editors"
memory(action="replace", target="memory",
       old_text="dark mode",
       content="User prefers light mode in VS Code, dark mode in terminal")

If the substring matches multiple entries, an error is returned asking for a more specific match.

memory — Agent's Personal Notes

For information the agent needs to remember about the environment, workflows, and lessons learned:

• Environment facts (OS, tools, project structure)
• Project conventions and configuration
• Tool quirks and workarounds discovered
• Completed task diary entries
• Skills and techniques that worked

user — User Profile

For information about the user's identity, preferences, and communication style:

• Name, role, timezone
• Communication preferences (concise vs detailed, format preferences)
• Pet peeves and things to avoid
• Workflow habits
• Technical skill level

Save These (Proactively)

The agent saves automatically — you don't need to ask. It saves when it learns:

• User preferences: "I prefer TypeScript over JavaScript" → save to user
• Environment facts: "This server runs Debian 12 with PostgreSQL 16" → save to memory
• Corrections: "Don't use sudo for Docker commands, user is in docker group" → save to memory
• Conventions: "Project uses tabs, 120-char line width, Google-style docstrings" → save to memory
• Completed work: "Migrated database from MySQL to PostgreSQL on 2026-01-15" → save to memory
• Explicit requests: "Remember that my API key rotation happens monthly" → save to memory

Skip These

• Trivial/obvious info: "User asked about Python" — too vague to be useful
• Easily re-discovered facts: "Python 3.12 supports f-string nesting" — can web search this
• Raw data dumps: Large code blocks, log files, data tables — too big for memory
• Session-specific ephemera: Temporary file paths, one-off debugging context
• Information already in context files: SOUL.md and AGENTS.md content

Memory has strict character limits to keep system prompts bounded:

What Happens When Memory is Full

When you try to add an entry that would exceed the limit, the tool returns an error:

{
  "success": false,
  "error": "Memory at 2,100/2,200 chars. Adding this entry (250 chars) would exceed the limit. Consolidate now: use 'replace' to merge overlapping entries into shorter ones or 'remove' stale or less important entries (see current_entries below), then retry this add — all in this turn.",
  "current_entries": ["..."],
  "usage": "2,100/2,200"
}

The agent should then:

1. Read the current entries (shown in the error response)
2. Identify entries that can be removed or consolidated
3. Use replace to merge related entries into shorter versions
4. Then add the new entry

Best practice: When memory is above 80% capacity (visible in the system prompt header), consolidate entries before adding new ones. For example, merge three separate "project uses X" entries into one comprehensive project description entry.

Practical Examples of Good Memory Entries

Compact, information-dense entries work best:

# Good: Packs multiple related facts
User runs macOS 14 Sonoma, uses Homebrew, has Docker Desktop and Podman. Shell: zsh with oh-my-zsh. Editor: VS Code with Vim keybindings.

# Good: Specific, actionable convention
Project ~/code/api uses Go 1.22, sqlc for DB queries, chi router. Run tests with 'make test'. CI via GitHub Actions.

# Good: Lesson learned with context
The staging server (10.0.1.50) needs SSH port 2222, not 22. Key is at ~/.ssh/staging_ed25519.

# Bad: Too vague
User has a project.

# Bad: Too verbose
On January 5th, 2026, the user asked me to look at their project which is
located at ~/code/api. I discovered it uses Go version 1.22 and...

The memory system automatically rejects exact duplicate entries. If you try to add content that already exists, it returns success with a "no duplicate added" message.

Memory entries are scanned for injection and exfiltration patterns before being accepted, since they're injected into the system prompt. Content matching threat patterns (prompt injection, credential exfiltration, SSH backdoors) or containing invisible Unicode characters is blocked.

Beyond MEMORY.md and USER.md, the agent can search its past conversations using the session_search tool:

• All CLI and messaging sessions are stored in SQLite (~/.hermes/state.db) with FTS5 full-text search
• Search queries return actual messages from the DB — no LLM summarization, no truncation
• The agent can find things it discussed weeks ago, even if they're not in its active memory
• The agent can also scroll forward/backward inside any session it finds

hermes sessions list    # Browse past sessions

See Session Search Tool for the three calling shapes (discovery / scroll / browse) and the response format.

sessionsearch vs memory

Memory is for critical facts that should always be in context. Session search is for "did we discuss X last week?" queries where the agent needs to recall specifics from past conversations.

The learning journey is a timeline view of everything Hermes has learned — saved skills and memory entries plotted over time (oldest at top, newest at bottom), with a playable "constellation" scrubber that replays the build-up. The same graph data drives three surfaces:

• Classic CLI / standalone — hermes journey (aliases: hermes learning, hermes memory-graph) renders the timeline in the terminal. Flags: --play animates the build-up (--fps to tune it), --width/--height override the render size, --no-color disables color, and --json dumps the raw graph payload.
• TUI — /journey (aliases: /learning, /memory-graph) opens the timeline as an overlay.
• Desktop app — /journey opens the Star Map / memory-graph panel, an interactive visual of the same nodes.

Beyond viewing, the journey is also where you prune and correct what Hermes has learned:

The same list / delete <id> / edit <id> subcommands work from the in-chat /journey command on the CLI, and the desktop panel offers edit/delete on nodes directly.

By default the agent saves memory freely — including from the background self-improvement review that runs after a turn. If you'd rather approve saves first, set memory.write_approval: true. It's a simple on/off gate applied to both foreground turns and the background review:

To turn memory off entirely (not just gate it), set memory_enabled: false.

Review staged writes from the CLI or any messaging platform:

/memory pending             # list staged memory writes (auto ones tagged [auto])
/memory approve <id>        # apply one (or 'all')
/memory reject <id>         # drop one (or 'all')
/memory approval on         # turn the gate on (or 'off') and persist it

This is the answer to "the agent saved a wrong assumption about me": set write_approval: true, and every save — especially the unprompted background ones — waits for your yes/no before it ever enters your profile.

After a turn, the background self-improvement review may quietly save a memory or update a skill. This is Hermes' consent-aware learning loop: repeated corrections and durable workflow lessons become compact memory entries or procedural skills, while write_approval can stage those writes for review before they affect future sessions. By default it surfaces a short 💾 Memory updated line in chat so you know it happened. Control how chatty that is:

display:
  memory_notifications: on    # off | on (default) | verbose

This only governs the gateway chat notification. The review itself, and writes to your memory/skill stores, are unaffected by this setting. Set it per-platform via display.platforms.<platform>.memory_notifications.

The review runs on your main chat model by default, replaying the conversation — which is already warm in the prompt cache, so it's cheap cache reads. On an expensive main model you can run the review on a cheaper model instead:

auxiliary:
  background_review:
    provider: openrouter
    model: google/gemini-3-flash-preview   # auto (default) = main chat model

When you point it at a model different from your main one, the review runs there for substantially lower cost (~3–5× in benchmarks). Because a different model can't reuse your main model's prompt cache anyway, the fork automatically replays a compact digest of the conversation (recent turns verbatim + a summary of older ones) rather than the full transcript — minimizing what it writes to the new cache. Capture holds: in testing, memory capture was identical and skill capture near-identical to the main-model review.

Leave it at auto (or set it to your main model) and nothing changes — the review keeps running on the main model with the full warm-cache replay.

Disabling automatic reviews (enabled)

The review fork can burn a meaningful share of total tokens on busy hosts. Operators can disable it without zeroing nudge intervals:

auxiliary:
  background_review:
    enabled: true              # false = skip automatic post-turn forks

With enabled: false, automatic post-turn forks do not spawn; manual /refine still works.

Fork usage is persisted in session_model_usage with task='background_review' and a completion line is written to agent.log (Background review complete: thread=bg-review calls=… in=… out=… result=…).

Skills use the same on/off gate, but the review UX differs because a SKILL.md is far too large to read in a chat bubble:

skills:
  write_approval: false     # false = write freely (default) | true = require approval

When write_approval: true, skill writes (create / edit / patch / write_file / delete) always stage regardless of origin. You review the one-line gist inline, but the full diff stays out-of-band:

/skills pending             # list staged skill writes + a one-line gist each
/skills diff <id>           # full unified diff (best viewed in CLI or dashboard)
/skills approve <id>        # apply it (or 'all')
/skills reject <id>         # drop it (or 'all')
/skills approval on         # turn the gate on (or 'off') and persist it

On a messaging platform, approve a skill from its gist + metadata, or open /skills diff on the CLI / dashboard / the staged file under ~/.hermes/pending/skills/<id>.json when you want to read the whole change. Full details in Gating agent skill writes.

For deeper, persistent memory that goes beyond MEMORY.md and USER.md, Hermes ships with 8 external memory provider plugins — including Honcho, OpenViking, Mem0, Hindsight, Holographic, RetainDB, ByteRover, and Supermemory.

External providers run alongside built-in memory (never replacing it) and add capabilities like knowledge graphs, semantic search, automatic fact extraction, and cross-session user modeling.

hermes memory setup      # pick a provider and configure it
hermes memory status     # check what's active

See the Memory Providers guide for full details on each provider, setup instructions, and comparison.