| name | hmem-update |
| description | Update flow for hmem. Runs `npm update -g`, syncs skills, applies migrations, verifies hooks, shows the changelog. Use when the user asks to update/upgrade hmem, hmem-mcp, or its-over-9k (any language), or when the startup version-check flags a new release. Runs the npm update itself — don't assume it's already done. |
/hmem-update — Update Routine
Drives the full update flow for hmem and hmem-sync: detects the current
version, runs npm update -g if outdated, syncs skill files, applies any
migrations, verifies hooks and configs, runs the smoke test. Every step is
important — do not skip steps.
Package naming note: The npm package is hmem (formerly its-over-9k,
before that hmem-mcp). The installed CLI is still hmem, the MCP server tools
still use the hmem prefix, and the GitHub repo is Bumblebiber/hmem.
The its-over-9k name now belongs to a separate project (the o9k efficiency
meta-framework) — do not npm-update via that name anymore.
Always use the current package name (hmem) for npm commands.
Version mapping: The renames reset the version line twice:
hmem-mcp (legacy) | its-over-9k (1.x line) | hmem (current) |
|---|
| up to 7.4.x | 1.0.0 – 1.3.8 | 8.0.0+ |
Steps 2d–2l below describe migrations only relevant when upgrading from the
legacy 5.x/6.x/7.x line. Skip them if your installed version is 1.x or 8.x —
they are already baked into the rename baseline.
Upgrading from its-over-9k 1.x: uninstall the old package and install the
new name — npm update -g its-over-9k will never reach 8.0:
npm uninstall -g its-over-9k
npm install -g hmem
npx hmem update-skills
The skills were renamed o9k-* → hmem-* in 8.0.0. update-skills cleans up
stale o9k-* folders automatically.
Step 1: Version Check
Determine the current and latest version:
hmem --version
npm view hmem version
npm view hmem versions --json
Read the changelog for the version range:
cd ~/projects/hmem && git log --oneline <old-tag>..HEAD
Or check GitHub releases: gh release list -R Bumblebiber/hmem --limit 5
If already on latest: Tell the user and skip to Step 7 (smoke test).
Step 1b: Run the Update
If the installed version is older than the latest on npm, run the update now:
npm update -g hmem
If the user installed via a deprecated package name (its-over-9k or hmem-mcp),
uninstall it first to avoid two parallel installs:
npm list -g --depth=0 | grep -E 'hmem-mcp|its-over-9k|hmem'
npm uninstall -g its-over-9k
npm uninstall -g hmem-mcp
npm install -g hmem
After the update, the MCP server is still running the OLD version —
it's loaded into the host process (Claude Code, Gemini CLI, etc.) and
won't pick up the new code until restart. Continue with Steps 2–6 first;
the restart happens in Step 7.
Step 1c: Present the changelog to the user
Always — even for a single-patch jump. Right after the update is installed, fetch the GitHub release notes for every version in the gap (<old> exclusive → <new> inclusive) and present a tidy summary to the user. They just installed something; they deserve to know what it does.
OLD=<previously-installed version>
NEW=$(node -p "require('/path/to/hmem/package.json').version")
gh release list -R Bumblebiber/hmem --limit 20 \
| awk -v o="v$OLD" -v n="v$NEW" '$1 > o && $1 <= n {print $1}' \
| while read tag; do
echo "=== $tag ==="
gh release view "$tag" -R Bumblebiber/hmem --json body -q .body
done
Present to the user in this shape (German if user speaks German, English otherwise):
📦 hmem v1.2.3 → v1.2.5 — Was sich geändert hat:
v1.2.5 — Pull from hmem-sync before first-message context
• Hook-startup pulls latest entries before building greeting/project list
• New /hmem-release rule: release notes for every release
v1.2.4 — Natural session-start greeting
• One-line greeting with name + 🟢/🟡/🔴 sync dot
• Conditional project list if no project named in opening message
v1.2.3 — hmem-sync link status in session-start context
• New --- hmem-sync --- block shows whether writes propagate
Keep each release to 2–4 bullets. Skip the "Impact / Migration" section unless it requires user action — flag those separately as 🛠 actions after the summary.
If gh is not available or the user is offline, fall back to:
cd ~/projects/hmem && git log --oneline "v$OLD..v$NEW"
and present commit titles instead.
Step 2: Update Skills
hmem update-skills
This syncs all skill files from the npm package to the local skills directory
and prunes stale hmem-* skills that are no longer bundled (as of v6.3.2).
Example output: × hmem-self-curate (removed, no longer bundled).
Verify:
ls ~/.claude/skills/hmem-*/SKILL.md
ls ~/.config/gemini/skills/hmem-*/
Check for new skills that weren't there before — inform the user about new capabilities.
If a skill was removed (e.g. merged into another), mention that too so the user knows
the workflow has moved.
Step 2b: Verify Hooks
Hooks are critical — without them, O-entries are never logged and auto-checkpoints never fire.
Check the current hook configuration. Use the platform-appropriate command:
cat ~/.claude/settings.json | grep -A5 hooks
# Windows (PowerShell)
Get-Content "$env:USERPROFILE\.claude\settings.json" | Select-String -Pattern "hooks" -Context 0,5
Required hooks (for checkpointMode: "auto"):
- UserPromptSubmit — memory load + checkpoint reminder
- Stop — exchange logging (
hmem log-exchange) + O-entry title generation
- SessionStart[clear] — context re-injection after
/clear
If hooks are missing or empty (hooks: {}):
- Inform the user: "Hooks are not configured — O-entries won't be logged and auto-checkpoints won't fire."
- Suggest: "Run
/hmem-config to set up hooks, or run hmem init to re-initialize."
If hooks exist but reference old paths or scripts:
- Check that hook scripts exist and are executable
- Verify they reference the current hmem installation path
Windows-specific hook checks (CRITICAL)
On Windows, two specific issues break hooks. Always run these checks when updating on Windows:
Check 1 — shell: powershell present on every hook command?
Each object in hooks.*.hooks and the statusLine object must contain "shell": "powershell". Without it, Claude Code may route the command through Git Bash, whose MSYS2 runtime crashes transiently at startup (bash.exe: *** fatal error - add_item ... errno 1) before the command is even parsed. Every hook then fails with a generic error.
Check 2 — No inline env-var syntax in commands?
Commands must NOT contain VAR=value prefixes like HMEM_PATH=C:/... node .... That's bash-only syntax; cmd.exe and PowerShell interpret HMEM_PATH=... as the command name and fail. All env vars must live in the top-level env block of settings.json.
The correct Windows shape:
{
"env": {
"HMEM_PATH": "C:/Users/<you>/.hmem/Agents/<AGENT>/<AGENT>.hmem"
},
"hooks": {
"Stop": [
{
"matcher": "",
"hooks": [
{
"type": "command",
"command": "node C:/Users/<you>/AppData/Roaming/npm/node_modules/hmem/dist/cli.js log-exchange",
"shell": "powershell"
}
]
}
]
}
}
If either check fails: Offer to fix settings.json automatically. The fix is lossless on other platforms, so it's safe to apply even on shared configs synced across OSes. Point the user to the Windows hook section in /hmem-config for the full pattern (UserPromptSubmit, Stop, SessionStart, statusLine).
After fixing: Claude Code must be restarted so the env block is re-loaded and hooks are re-registered with the new shell.
Step 2c: Check load_project Display Config
Since v5.1.8, load_project supports configurable section expansion:
loadProjectExpand.withBody: sections showing L3 title + body (default: [1] = Overview)loadProjectExpand.withChildren: sections listing all L3 children as titles (default: [6, 8] = Bugs, Open Tasks)
Check if the user has customized this in hmem.config.json. If not, inform them about the option:
{ "memory": { "loadProjectExpand": { "withBody": [1], "withChildren": [6, 8, 10, 16] } } }
Section 16 = Rules — include it so project-specific agent directives are visible on every load_project.
Legacy migration steps (Step 2d–2l)
The following sub-steps apply only when upgrading from the pre-rename hmem-mcp
line (≤ 7.4.x). If hmem --version reports a 1.x version, the migrations are
already applied at install time — skip directly to Step 3.
Step 2d: HMEM_PATH Migration (v6.0.0+)
v6.0.0 replaced HMEM_PROJECT_DIR + HMEM_AGENT_ID with a single HMEM_PATH env var.
Check if migration is needed:
- Look at the user's
.mcp.json or ~/.claude.json for hmem env vars
- If you see
HMEM_PROJECT_DIR and/or HMEM_AGENT_ID → migration needed
Migration steps:
-
Determine the current .hmem file path:
- With agent ID:
{HMEM_PROJECT_DIR}/Agents/{HMEM_AGENT_ID}/{HMEM_AGENT_ID}.hmem
- Without:
{HMEM_PROJECT_DIR}/memory.hmem
-
Update MCP config — replace the old env vars with HMEM_PATH:
{
"env": {
"HMEM_PATH": "/absolute/path/to/your/file.hmem"
}
}
Remove HMEM_PROJECT_DIR, HMEM_AGENT_ID, and HMEM_AGENT_ROLE from the env block.
-
The .hmem file does NOT need to move — HMEM_PATH points to it wherever it is.
-
If hmem-sync is installed, also update to v1.0.0+ (npm update -g hmem-sync).
The --agent-id flag was removed — use --hmem-path or HMEM_PATH instead.
-
CRITICAL — Sync filename must match across all devices:
hmem-sync identifies stores by the local filename (e.g. DEVELOPER.hmem). If Device A
syncs as DEVELOPER.hmem and Device B syncs as memory.hmem, they will NOT see each
other's data — the server treats them as separate stores.
Check: Run hmem-sync status on each device. The "hmem file" line shows the filename
that will be used for sync. All devices sharing the same memory MUST use the same filename.
Common mistake after v6.0 migration: Devices that used HMEM_AGENT_ID=DEVELOPER
have DEVELOPER.hmem. New devices default to memory.hmem. These won't sync.
Fix: Rename the .hmem file on the mismatched device:
mv ~/.hmem/memory.hmem ~/.hmem/DEVELOPER.hmem
Then update HMEM_PATH in the MCP config to point to the renamed file.
Also removed in v6.0.0:
min_role parameter from write_memory and update_memory tools- Company store role gating (all agents can now write to company store)
HMEM_AGENT_ROLE / COUNCIL_AGENT_ROLE env vars
Step 2e: v7.0.0 — Two-Server Split
Only needed when upgrading from < v7.0.0
v7.0.0 moves 11 curation/maintenance tools into a separate hmem-curate binary.
The daily server now only exposes the 11 daily-use tools — less context noise, no accidental curation.
Tools that moved to hmem-curate: update_many, export_memory, import_memory,
memory_stats, memory_health, tag_bulk, tag_rename, move_memory, rename_id, move_nodes
Note (v1.3.0+): reset_memory_cache is now available in both the main server and hmem-curate — no curate server needed to reset the session cache.
1. Add hmem-curate to MCP config
Open ~/.mcp.json (or wherever hmem is configured) and add a second entry:
"hmem-curate": {
"command": "hmem-curate",
"env": {
"HMEM_PATH": "/path/to/your/file.hmem",
"HMEM_SYNC_PASSPHRASE": "..."
}
}
Leave it disabled by default — only activate via /mcp when running /hmem-curate or /hmem-migrate-o.
Verify the binary installed correctly:
hmem-curate --version
If hmem-curate is not found: npm update -g hmem (postinstall sometimes skips bin links).
2. Set up the hmem-using-hmem session hook
v7.0.0 ships a meta-skill that injects dispatch/memory habits at session start — similar to Superpowers' using-superpowers. One command does everything:
hmem setup-hook
This copies scripts/hmem-session-inject.sh to ~/.claude/hooks/ and registers the SessionStart hook in settings.json. Idempotent — safe to run again if already installed.
After the next session restart, hmem-using-hmem meta-rules will be in context automatically.
Step 2f: v7.0.0 — Device & Rate Limits in Statusline
Only needed when upgrading from < v7.0.0
1. Active Device in Statusline
The statusline now shows which device the session is running on (using an I-entry):
Strato Server | P0048 hmem | 3/5 | 5h: 34%/w: 17%
- Shows the I-entry title (e.g., "Strato Server"), no ID prefix
- Shows "identify device" in gray if not yet configured
- On first message of a new session, the agent auto-detects the device and calls
set_active_device
Set up the device (once per machine — persists across sessions):
set_active_device({ id: "I00XX" }) # use the I-entry ID for this machine
Device is stored in ~/.hmem/active-device. To verify:
cat ~/.hmem/active-device
If no I-entry exists for this machine yet: create one first with write_memory(prefix="I", ...).
2. Rate Limits in Statusline (Claude Max)
If you have a Claude Max subscription, the statusline automatically shows 5-hour and weekly usage:
5h: 34%/w: 17%
Colors: green (<50%), yellow (50–79%), red (≥80%). No config needed — data comes from Claude Code automatically. Only visible if Claude Code passes the data (Claude Max subscribers only).
Step 2g: v7.0.2 — Schema-Section Reconcile for append_memory
Only needed when upgrading from < v7.0.2
append_memory to a schema-enforced root entry (e.g. append_memory(id="I0002", ...)) is now
allowed if the content's first line matches a defined schema section name. Previously this was
always blocked ("uses a fixed schema — cannot add new L2 nodes directly").
Use case: Adding a section that was added to hmem.config.json after the entry was created
(e.g., adding the new Rules section to existing I-entries). No manual migration needed —
just run the reconcile step from Step 4c.
No schema changes, no DB migrations.
Step 2h: v7.0.4 - Null-Safe Titles and Richer Session-Start Context
Only needed when upgrading from < v7.0.4
Null-safe title fixes (set_active_device + statusline)
Some I-entries (and potentially other entries) have a NULL title in SQLite. In prior versions this caused set_active_device and the statusline to crash with Cannot read properties of null (reading split). Both are now fixed with (row.title ?? id).split("|")[0].
No action needed - the fix is purely in the code. If set_active_device was crashing before, it will now work correctly.
hook-startup enhancements (first-message context injection)
The UserPromptSubmit hook now injects richer context on the first message of every session:
-
Always-on device reminder - regardless of whether a device is already set, the agent is reminded to verify the active device matches the current machine and call set_active_device if needed. Previously this reminder was only shown when no device was set at all.
-
H-entries (human profile) - up to 10 H-entries (sorted by access count) are injected as a short list (ID title). Gives the agent immediate context about the user without requiring a separate read_memory(prefix="H") call.
-
Recent projects - the 3 most recently updated P-entries are injected as ID title. Helps the agent recognize which project to load if the user's first message is project-related.
No config change needed - all enhancements are automatic from the updated cli-hook-startup.js.
Windows example config on GitHub
settings.windows.example.json is now in the Bumblebiber/hmem repository. If you are on Windows, refer to it as the canonical example for hook and statusline configuration with "shell": "powershell".
Step 2i: v7.1.0 — load_project Noise Reduction + Skill Improvements
Only needed when upgrading from < v7.1.0
load_project: DONE-filter in schema mode
load_project now hides ✓ DONE and DONE-prefixed items from Roadmap and Next Steps sections when using schema-driven rendering. Previously this filter only applied to legacy rendering mode (no schema). No config change needed — automatic.
load_project: Project-specific R-entries only
Rules (R-prefix entries) are now only shown in load_project if the rule has an explicit link to the current project (r.links.includes(projectId)). Global rules without a project link are no longer injected. This eliminates ~1k token duplication for projects with many global rules. Action: If you have project-specific rules that should appear in load_project, ensure they have a links entry pointing to the project.
Skill improvements
- hmem-dispatch: New optional
VERIFY field — specify a shell command the sub-agent must run and return as [VERIFY_RESULT]. Use for code-change tasks to close the feedback loop.
- hmem-curate: Health check findings now classified as BLOCKER / WARNING / INFO for structured triage.
- hmem-write: New "After Writing" verification step — for complex L/D/E entries, read back the written entry and verify it captures "why?", not just "what?".
Step 2j: v7.2.0 — Codebase Node Schema + checkpointPolicy Fix
Only needed when upgrading from < v7.2.0
checkpointPolicy: readonly now correctly scoped
readonly on a schema section previously blocked ALL append_memory calls under that section (including adding L3 modules or L4 function signatures). This was a bug — the intent was only to prevent adding new L2 sections to a project root.
New behavior:
append_memory(id="P0048") → blocked if schema has readonly sections (prevents new L2 sections) ✓append_memory(id="P0048.2") → allowed (add module to Codebase) ✓append_memory(id="P0048.2.N") → allowed (add function signature) ✓pointer policy still enforced at all depths ✓
Action: Change "checkpointPolicy": "readonly" on the Codebase section in your hmem.config.json to "checkpointPolicy": "append" — the section is now freely appendable by agents.
Codebase node schema standardized
The .2 Codebase section now has a standard 4-level structure:
- L3 — Pipeline (auto-created via
defaultChildren): data flow overview
- L3.N — Module: title = filename, body = purpose +
src/file.ts
- L4 — Function (mandatory): title = full TS signature, body = description +
src/file.ts
- L5 — Extended Notes (optional): usage example, caveats, complex param details
Agents update L4 after every code change (enforced via hmem-subagent POST-TASK NODE SYNC).
Action: Run load_project on each active project — the Pipeline sub-node is auto-created via defaultChildren reconcile if missing.
Step 2k: v7.2.2 — Richer Session-Start Context + I-Entry Active Device Block
Only needed when upgrading from < v7.2.2
context-inject: 5 most recent projects instead of full list
hmem context-inject (SessionStart[clear] hook) now shows only the 5 most recently edited P-entries instead of all projects. A hint is appended:
(full list: read_memory({prefix:"P", titles_only:true}))
No config change needed — automatic.
hook-startup: H-entry title fix
H-entries without a title field (entries whose content lives entirely in level_1) now display correctly. Previously they appeared as blank lines in the --- Human context (H-entries) --- block.
hook-startup: Active device block
The first-message context injection now includes a --- Active device (I00XX) --- block showing:
- I-entry body (L1) — device name, specs, IP
- All L2 section titles (Specs, OS, IP Address, Access, Services, Apps, Storage, Notes, Rules)
- For the Apps section: full L3 title list (installed software)
- Irrelevant L2/L3 nodes are filtered out
hook-startup: recent projects limit raised 3 → 5
The --- Recent projects --- block in the first-message injection now shows 5 most recently updated P-entries (was 3).
Step 2t: v1.3.8 — Pre-injected session context + checkpointMode fix
Only relevant for Hermes Agent users.
hmem-startup.sh: stable context pre-injected — no more MCP roundtrip
The hook now calls hmem hook-startup (with explicit stdin pipe) on the first turn and
injects ~3.7k chars of stable context directly: H-entries, recent projects, device info,
sync status, infrastructure favorites, checkpoint status. The hmem-session-start skill
starts at STEP 1 (project activation) — the expensive read_memory(mode='essentials')
call is eliminated. Session-start time drops from ~20s to ~4s.
hmem-startup.sh: checkpointMode respected
The hook now reads checkpointMode from hmem.config.json and only shows checkpoint
reminders when set to "remind". In "auto" mode (default), reminders are suppressed —
checkpoints fire silently in the background.
hmem-session-start skill: STEP 0 removed, H-node detection expanded
- STEP 0 (
read_memory(mode='essentials')) removed — context is pre-injected
- STEP 5b expanded: full table of 7 required H-slots with detection guide (empty node =
no
• children), procedure for asking the user when H0008 is empty, and opportunistic
fill for other slots
Action (Hermes hosts):
cp ~/projects/hmem/hermes-hooks/hmem-startup.sh ~/.hermes/agent-hooks/hmem-startup.sh
chmod +x ~/.hermes/agent-hooks/hmem-startup.sh
(or keep the symlink — no redeploy needed)
No TypeScript changes. No npm behavior change. Hook scripts + skill only.
Only relevant for Hermes Agent users.
hmem-startup.sh rewritten — minimal bootstrap, skill-driven session start
The previous Hermes hmem-startup.sh duplicated session-start logic (H-entries, projects, sync status, greeting) that the hmem-session-start skill already handles better. The new version is a thin shim: sync-pull, write session-id cache for the statusline, inject a one-line directive that invokes the skill, then let the skill do everything else. Subagent calls now silent-pass-through. Long-session warnings added.
Action (Hermes hosts):
cp ~/projects/hmem/hermes-hooks/hmem-startup.sh ~/.hermes/agent-hooks/hmem-startup.sh
chmod +x ~/.hermes/agent-hooks/hmem-startup.sh
(or use the symlink option below to never deal with this again.)
Symlink deployment option
hmem-release Step 7b now documents two deployment modes for hermes-hooks/:
- Copy (default — explicit redeploy each update)
- Symlink (
ln -sf ~/projects/hmem/hermes-hooks/$f ~/.hermes/agent-hooks/$f per file — repo changes are instantly live, no redeploy needed)
Symlinks are recommended on dev devices where the repo lives in a stable path. End users on npm install -g hmem can symlink to node_modules/hmem/hermes-hooks/ if their global install path is stable across updates.
Step 2r: v1.3.6 — Leaner session-start + load_project drill IDs + per-device skill opt-out
Only relevant when upgrading from ≤ v1.3.5.
Rules block (post-/clear injection) now filters to pinned/favorite only
The ## Rules: listing emitted by the SessionStart:clear hook used to dump every non-obsolete R-entry — easily a dozen lines of always-on context. It now only shows R-entries with pinned=true or favorite=true.
Action: If a rule must remain omnipresent in every session, pin it explicitly:
update_memory(id="R00XX", pinned=true)
On first run after upgrade, only previously-pinned rules (typically R0027, R0026, R0010 in a default setup) will surface.
First-message directive shortened
The 10-line "STEP 1 / STEP 2 / STEP 3" block in UserPromptSubmit is now a single sentence (with a hasIntent branch). Saves ~5k tokens on every first message. The hmem-session-start skill still drives the actual session-start workflow.
load_project body sanitisation
XML fragments from tool-call output (</content>, <parameter ...>, <function_calls>) that occasionally leak into stored memory are now stripped on display. Bodies in your hmem are not modified — only the rendered output is cleaned. If you previously saw garbled L2/L3 bodies in load_project output, they should now render cleanly.
Truncation markers include drill IDs
[+N] cliffhangers in load_project are now [+N → read PXXXX.Y.Z] so the agent can drill directly without guessing the path.
/hmem-new-rule skill added
Use it whenever adding a rule:
/hmem-new-rule
It enforces the cross-project (R-prefix) vs. project-specific (subnode under active project's Rules section) decision before writing — a check that the agent has gotten wrong before.
Per-device skill exclusion via ~/.hmem/skills-disabled
If certain bundled skills are clutter on a specific device (e.g. setup/migration skills you've already run), create ~/.hmem/skills-disabled with one skill name per line:
hmem-setup
hmem-migrate-o
hmem-sync-setup
hmem update-skills now skips them and removes any previously-installed copies from ~/.claude/skills/. The repo continues to ship all skills; only this device opts out. Lines starting with # are comments.
hmem-release skill covers hook artifacts
The release checklist (Step 7b + Quick Reference) now explicitly audits hermes-hooks/, plugins/hermes-hmem/, and scripts/hmem-*.sh on every release. Previously, hook changes shipped silently with no verification step.
Step 2l: v7.4.0 — syncSecrets default flipped + non-blocking sync
Only needed when upgrading from < v7.4.0
syncSecrets default is now false (was true) — BREAKING
Sync configs that omitted syncSecrets previously pushed tokens/salts to the sync server. After upgrading, secrets stay local unless "syncSecrets": true is set explicitly in hmem.config.json. Motivated by the 2026-05-05 credential-exposure incident.
Action:
- If you rely on secret-sync between devices (e.g. for
hmem-sync restore shortcuts), add "syncSecrets": true to your sync block.
- Otherwise: nothing to do. The safer default kicks in automatically.
Check your config:
grep -A 10 '"sync"' "$(dirname "$HMEM_PATH")/hmem.config.json"
Non-blocking sync I/O
syncPull, syncPullThenPush, syncPushSync, syncPushWithRetry, reserveId, reserveNextId, and reserveNextSubIds are now async. The stdio transport is no longer frozen during a slow push. Tool behaviour is unchanged for callers.
Internal: HmemStore.db is @internal
If you wrote third-party code that called store.db.prepare(...), prefer the new public methods (isObsolete, hasActiveEntryWithPrefix, getNonObsoleteTitle) — direct access still works but bypasses the integrity-check guard.
Internal: migration tracking via schema_version
The MIGRATIONS array of ALTER TABLE statements is now tracked in schema_version (alter_vN). Genuine errors are logged instead of silently swallowed. No action needed; first open after upgrade marks all known migrations as applied.
Step 2q: Hermes Agent Hooks Deployment
Only relevant for Hermes Agent users.
Since v1.3.1, the repo ships hermes-hooks/ with scripts for Hermes Agent
statusline integration. After npm update, deploy them:
cp ~/projects/hmem/hermes-hooks/*.sh ~/.hermes/agent-hooks/
chmod +x ~/.hermes/agent-hooks/*.sh
The hooks provide:
hmem-startup.sh — Sync-Status, Projekt-Liste, session_id-Cache (pre_llm_call)hmem-log-exchange.sh — Exchange-Logging an hmem (post_llm_call)hmem-statusline.sh — CC-Style Statusbar: Device | Projekt → O-Node | Counter
Hermes CLI Patch: The statusline integration also requires a small patch
to Hermes Agent's cli.py. Apply with:
cd ~/.hermes/hermes-agent
git apply ~/projects/hmem/hermes-hooks/hermes-cli-hmem-statusline.patch
Hermes config.yaml must reference the hooks (see /hmem-config for the
full configuration template). After applying the patch and deploying hooks,
restart Hermes for changes to take effect.
Step 2p: v1.3.0 — reset_memory_cache in main server + harness-aware checkpoint
Only relevant when upgrading from < v1.3.0
reset_memory_cache now in main server — call it directly without needing the hmem-curate server. Useful when load_project returns "already active (loaded recently)" and you need a forced reload.- Harness-aware checkpoint routing — Claude Code → Haiku via
claude -p (Max OAuth), Codex → gpt-5.4-mini, others → configured provider. No config change needed.
- Orphan-batch catchup — checkpoint agent now processes up to 5 unsummarized past batches per run. Run
hmem checkpoint manually to catch up a backlog.
- Pi hooks known issue — hook-based exchange logging is currently broken in the Pi runtime. Workaround: run
hmem checkpoint manually.
No migration required.
Step 2o: v1.2.5 — Pull from hmem-sync before first-message context
Only relevant when upgrading from < v1.2.5
hmem hook-startup now calls syncPull(HMEM_PATH) before reading local SQLite for the greeting/project list. Closes the gap where entries written on another device weren't visible until the agent's first MCP read_memory call triggered the existing pull.
- Bounded to 3 seconds — falls back to stale local data if the server is slow or unreachable
- No-op if hmem-sync isn't configured / no
HMEM_SYNC_PASSPHRASE env var
- 30s cooldown is per-process and won't double-pull with MCP
Typical added latency: ~0ms (no sync) to ~1s (cross-region pull). No action required.
Step 2n: v1.2.4 — Natural session-start greeting
Only relevant when upgrading from < v1.2.4
hmem hook-startup now drives a natural-language greeting on the first message of every session:
- Silent context load —
load_project if the user named one, else read_memory().
- One-line greeting in the user's preferred language (from H-entries), with name and a 🟢/🟡/🔴 sync dot.
- Conditional project list — if the first message did NOT mention a project (regex:
lade Projekt, load project, P\d{4}, aktiviere, wechsel zu, switch to project, work on P…, open project), the greeting is followed by the 5 most recent projects and a "Welches?" question.
- Then the agent handles the user's actual message.
Replaces the old [CORTEX READY] block output from hmem-session-start. No config needed — automatic on next session.
Step 2m: v1.2.3 — hmem-sync link status in session-start context
Only relevant when upgrading from < v1.2.3
hmem hook-startup now reads ~/.hmem/config.json on the first message of every session and appends a one-line --- hmem-sync --- block to the agent's additionalContext:
✓ Linked to <server> | active_file: <id> | last sync: <ago> — writes propagate to other devices on next hmem-sync push⚠ Linked … never synced — run hmem-sync pull to fetch⚠ Authenticated … no active file — run hmem-sync setup✗ Not linked — writes stay local; run hmem-sync login- Silent if
config.json doesn't exist (hmem-sync never configured)
The hmem-session-start skill maps this block to a visible 🟢/🟡/🔴 indicator in the [CORTEX READY] output, so the user sees connection state at a glance.
No action required — automatic on next session start.
Step 3: Entry Migration
Some versions introduce new data formats. Check if migration is needed:
v5.1.0+ Title/Body Separation:
v5.1.2+ Checkpoint Summaries:
v5.1.2+ Skill-Dialog Tags:
- Exchanges containing skill activations should be tagged
#skill-dialog
- These are auto-tagged by the checkpoint process going forward
- For old exchanges: the checkpoint auto-tagger picks them up on the next run
General migration pattern:
- Read a sample of entries to assess the current state
- Identify entries that don't match the new format
- Fix in batches — don't try to fix everything at once
- Prioritize: favorites and pinned entries first, then high-access, then the rest
Step 4: P-Entry Schema Enforcement (R0009)
All P-entries (projects) must follow the standard 16-section L2 structure defined in hmem.config.json. Core sections:
.1 Overview (readonly)
.2 Codebase (append — L3 modules + L4 signatures freely appendable by agents)
.3 Dependencies (readonly)
.4 Usage (readonly)
.5 Requirements (readonly)
.6 Context (readonly)
.7 Deployment (readonly)
.8 Security (pointer — only E-entry refs)
.9 Performance (pointer — only E-entry refs)
.10 Bugs (pointer — only E-entry refs)
.11 History (readonly — session log, chronological)
.12 Roadmap (append)
.13 Ideas (append)
.14 Team (readonly)
.15 Next Steps (append)
.16 Rules (readonly — project-specific agent directives)
checkpointPolicy controls what the Haiku checkpoint agent may write:
readonly — Haiku never modifies this sectionpointer — Haiku may only add nodes that reference an entry ID (e.g. [E0124])append — Haiku may freely add sub-nodes
Codebase node structure (.2):
- L3 first child = Pipeline (data flow overview)
- L3.N = Module (title: filename, body: purpose +
src/file.ts)
- L4 = Function (title: full TS signature, body: one-line description +
src/file.ts)
Agents update .2 after every code change (POST-TASK NODE SYNC in hmem-subagent skill).
For each active P-entry:
read_memory(id="P00XX", depth=2) — check L2 structure- Run
load_project(id="P00XX") — auto-reconcile adds any missing sections
- L1 body should be:
Name | Status | Stack | Repo
Do not restructure entries that already follow the schema. Only fix what's missing or wrong.
Step 4b: Config & DB Migration (Protocol → History + Rules)
This step is only needed when upgrading from a config that still has "Protocol" as a section name.
Check if migration is needed:
grep -c '"Protocol"' ~/.hmem/*/hmem.config.json ~/.hmem/Agents/*/hmem.config.json 2>/dev/null
If the output is 0 everywhere — skip this step.
1. Update hmem.config.json
In memory.schemas.P.sections, make these changes:
- Rename
"Protocol" → "History" and set "checkpointPolicy": "readonly"
- Add
"Rules" section at the end: { "name": "Rules", "loadDepth": 1, "checkpointPolicy": "readonly" }
- Add
checkpointPolicy to all sections per the table in Step 4
- In
loadProjectExpand.withChildren, add 16 for Rules visibility
Full recommended policies (add to each section):
{ "name": "Overview", "checkpointPolicy": "readonly" },
{ "name": "Codebase", "checkpointPolicy": "append" },
{ "name": "Dependencies", "checkpointPolicy": "readonly" },
{ "name": "Usage", "checkpointPolicy": "readonly" },
{ "name": "Requirements", "checkpointPolicy": "readonly" },
{ "name": "Context", "checkpointPolicy": "readonly" },
{ "name": "Deployment", "checkpointPolicy": "readonly" },
{ "name": "Security", "checkpointPolicy": "pointer" },
{ "name": "Performance", "checkpointPolicy": "pointer" },
{ "name": "Bugs", "checkpointPolicy": "pointer" },
{ "name": "History", "checkpointPolicy": "readonly" },
{ "name": "Roadmap", "checkpointPolicy": "append" },
{ "name": "Ideas", "checkpointPolicy": "append" },
{ "name": "Team", "checkpointPolicy": "readonly" },
{ "name": "Next Steps", "checkpointPolicy": "append" },
{ "name": "Rules", "loadDepth": 1, "checkpointPolicy": "readonly" }
2. Rename Protocol → History in existing P-entries
Find all Protocol section nodes and rename them:
sqlite3 /path/to/your.hmem \
"UPDATE memory_nodes SET title='History', content=REPLACE(content,'Protocol','History'), updated_at=datetime('now') WHERE LOWER(title)='protocol' AND root_id LIKE 'P%';
SELECT changes() || ' nodes renamed';"
Then rebuild the FTS index:
sqlite3 /path/to/your.hmem "PRAGMA wal_checkpoint(TRUNCATE);"
3. R-entry curation
Global R-entries should only contain rules that apply to ALL projects. Project-specific rules belong in the P-entry's Rules section.
Review all R-entries (read_memory(prefix="R")). For each one:
- Ask: "Would an agent working on an unrelated project need this?"
- If NO → move to the relevant P-entry's Rules section and mark the R-entry as irrelevant
Common candidates for migration:
- Release/publish rules → into the project's P-entry Rules
- Game/domain-specific constraints → into the relevant project's Rules
- Tool-specific workarounds → into the project's Rules
To add a rule to a P-entry:
append_memory(id="P00XX.YY", content="Rule text here (→ was R00ZZ)")
where P00XX.YY is the Rules section node ID (found via load_project).
Step 4c: I-Entry Rules Node Reconciliation
I-entries (Infrastructure/Devices) have a Rules section in the schema for device-specific
agent directives. When this section is added to hmem.config.json, existing I-entries do NOT
get it automatically — unlike P-entries (which reconcile on load_project), I-entries have
no auto-reconcile trigger.
Check which I-entries are missing Rules:
read_memory(prefix="I", depth=2)
Look for entries that don't show a Rules L2 node.
Add the missing Rules node (requires hmem-mcp ≥ v7.0.2 which allows schema-section
appends to existing root entries):
append_memory(id="I00XX", content="Rules\n\tDevice-specific directives here")
Fill in relevant directives per device — examples:
- Server: sudo access, package manager quirks, which user Claude runs as
- Dev machine: OS-specific conventions (apt vs dnf), primary project
- Mobile/laptop: availability constraints, work vs personal restrictions
- Services (npm, rmapi): access method, credentials location, rate limits
Note: If append_memory returns "uses a fixed schema — cannot add new L2 nodes directly"
even with "Rules" as the first line, the running MCP server is older than v7.0.2. Update
first (npm update -g hmem), then restart Claude Code, then add the Rules nodes.
Step 5: O-Entry Curation
Check recent O-entries for quality:
read_memory(prefix="O")
Titles:
- Replace "unassigned" or generic titles (e.g., "hmem-mcp") with descriptive ones
- Good: "Title/Body Separation design + v5.1.0 release"
- Fix:
update_memory(id="O00XX", content="Descriptive session title")
Tags:
- Every O-entry should have at least
#session
- Add topic tags where obvious:
#release, #bugfix, #refactor, #brainstorming
- Fix:
update_memory(id="O00XX", tags=["#session", "#release"])
Checkpoint Summaries:
- O-entries with >10 exchanges and no
[CP] summary need one
- Write summary:
append_memory(id="O00XX", content="\t[CP] Summary...")
- The auto-tagger will tag it
#checkpoint-summary on the next checkpoint run
Cleanup:
- Look for duplicate O-entries (same title, same date, 1-2 exchanges) — these are likely subagent artifacts
- Mark as irrelevant or delete if clearly junk
Step 6: hmem-sync Update (if installed)
Check if hmem-sync is installed and needs updating:
which hmem-sync && hmem-sync --version
npm view hmem-sync version
If outdated:
npm update -g hmem-sync
Verify sync still works:
hmem-sync status
hmem-sync push
hmem-sync pull
If hmem-sync is not installed: Skip this step. Mention to the user that hmem-sync is available for cross-device sync.
Step 7: Restart Prompt
IMPORTANT: The smoke test must run against the NEW MCP server version. Since the MCP
server is loaded into the host process (Claude Code, Gemini CLI, etc.), an npm update does
NOT take effect until the tool is restarted.
Tell the user:
All migration steps complete. Please restart Claude Code now to load the new MCP server.
After restart, run /hmem-update again — I'll skip straight to the smoke test.
If already on latest version (detected in Step 1): Skip this step — the MCP server
is already running the current version. Proceed directly to the smoke test.
After restart: When /hmem-update runs again and Step 1 shows "already on latest",
proceed to the smoke test immediately.
Step 8: Smoke Test
Verify everything works after the update. Only run this after the restart (or if no
update was installed — i.e., already on latest version).
read_memory() # bulk read works
read_memory(id="P00XX") # drill-down works
load_project(id="P00XX") # project loading works
read_project(id="P00XX") # v7.0.0: read without activating O-entry routing
write_memory(prefix="T", content="Update smoke test — delete me", tags=["#test"])
# write works → note the ID
update_memory(id="T00XX", content="Update smoke test — verified", irrelevant=true)
# update works + mark for cleanup
If any step fails: report the error to the user. Do not proceed with normal work until the issue is resolved.
Step 9: Report
Tell the user what was done. Always remind to restart if an actual update was
installed and the user hasn't restarted yet.
hmem updated: v1.1.0 → v1.2.1
Changes applied:
- Skills synced (2 new, 3 updated)
- 5 P-entries checked against R0009 schema (2 fixed)
- 12 O-entries curated (4 titles fixed, 3 summaries added)
- Smoke test passed ✓
Step 10: Surface Any Problems as GitHub Issues
If anything friction-y happened during this update — wrong package name in
the skill, broken migration step, confusing version mismatch, stale MCP config
that wasn't detected, a tool returning unexpected output, anything — tell the
user to file it (or offer to draft the issue yourself):
👉 https://github.com/Bumblebiber/hmem/issues
A good issue includes: installed version (hmem --version), OS, the exact
command that failed and its output, and what you expected. The faster these
land in the tracker, the faster the next release fixes them.
Also worth running once after the update:
hmem doctor
This scans ~/.claude.json for stale or deprecated hmem MCP entries (e.g. paths
from another device that no longer exist on this one, or env vars left over
from the pre-v6.0 HMEM_PROJECT_DIR + HMEM_AGENT_ID syntax). It reports
findings only — never auto-modifies host configs.
Auto-Detection (for hook integration)
This skill can be triggered automatically. At session startup, if the hmem MCP server detects that the installed version differs from the last-seen version stored in the config, it appends a notice to the first read_memory() response:
⚠ hmem updated: v1.1.0 → v1.2.1. Run /hmem-update to apply post-update steps.
The agent should then invoke this skill automatically or ask the user if they want to run it.
Last-seen version is stored in hmem.config.json under lastSeenVersion. Updated automatically after a successful /hmem-update run.
