hippo
v1.36.1Local, git-native memory for Claude Code — your repo is the store, recall is $0/prompt (no LLM, no network), staleness is git-drift (the cited code moved), and team memory ships through review.
By youknowfredLicense: MIT1 GitHub starsUpdated 2 hours ago
Directory evidence
- Runtimes
- Claude Code
- Parsed components
- 18 skill or MCP entries
- Source updated
- Oct 1, 2026
- Manifest status
- Canonical path parsed
The directory validates manifest shape and source location. It does not execute the plugin or provide a security endorsement. Review the indexing methodology →
Install hippo for Claude Code
claude plugin marketplace add IchenDEV/agent-plugin-mkt
claude plugin marketplace update agent-plugin-marketplace
claude plugin install hippo@agent-plugin-marketplacePaste and run these commands in a terminal with Claude Code. They add and refresh the PluginsMP catalog, then install this plugin.
The installer fetches third-party code from the source repository shown on this page. This directory validates manifest structure and source location, but does not perform a security audit; review the manifest, components, and source before installing.
Get the source manually
git clone https://github.com/youknowfred/hippoClone the source repository, then follow its setup instructions to add the plugin to a compatible client. The plugin root is plugin/.
Plugin files
├── .claude-plugin/plugin.json├── skills/bootstrap/SKILL.md├── skills/consolidate/SKILL.md├── skills/doctor/SKILL.md├── skills/dream/SKILL.md├── skills/export-agents/SKILL.md├── skills/import/SKILL.md├── skills/init/SKILL.md├── skills/new/SKILL.md├── skills/pack/SKILL.md├── skills/promote/SKILL.md├── skills/promote-rule/SKILL.md├── skills/publish/SKILL.md├── skills/recall/SKILL.md├── skills/remove/SKILL.md├── skills/resolve/SKILL.md├── skills/review/SKILL.md├── skills/why/SKILL.md└── .mcp.json
Included Skills17
Run this once per machine (not per project) to build the shared plugin venv and warm the offline embedding model cache. Use when a user says "bootstrap memory", "set up hippo", "/hippo:bootstrap", or when /hippo:doctor reports the venv/model cache is missing. Idempotent — safe to re-run; it no-ops via a sentinel file if already bootstrapped. Pass `--multilingual` to switch to a multilingual embedding model for non-English corpora.
Sleep-time consolidation — one deliberate turn for the write-side maintenance the hot path defers. Drains the CAP-2 pending-capture queue into memory (check-first, per-item approval), works the reconsolidation worklist (reverify/fix/demote/snooze), and refreshes the link graph. Triggers include "consolidate memory", "drain the capture queue", "process pending captures", "/hippo:consolidate". Suggested when the pending queue or stale worklist is deep. Not a content audit — that's /hippo:audit.
Fast health check for the memory plugin's own install/environment — is it bootstrapped, is the venv healthy, is the corpus symlinked and indexed correctly. Use for "is memory working", "check memory setup", "/hippo:doctor", or when recall seems to be silently returning nothing. This is a QUICK sanity check, not a deep corpus audit — for the latter use /hippo:audit.
The generative sleep pass — offline replay surfaces the latent graph edges consolidate can't reach (bridges, unlinked mentions, undeclared refines). Tier-A edges auto-apply reversibly; --dry-run previews; --deparasite runs the counterweight; reverse replay boosts ranking; --generate runs the quarantined DRM-6 schema/hypothesis draft tier (flagged, self-decaying, evidence-graduated). Triggers "dream", "find latent links", "de-parasite my graph", "generate schemas", "/hippo:dream".
Render this project's memory floor as a PROPOSED AGENTS.md diff — one ranked, drift-checked source for the cross-tool rule plane (Codex/Cursor/Copilot all read AGENTS.md). Glob scoping is derived from each memory's cited_paths; propose-only, never an authoritative overwrite; once applied, the exported file is drift-checked by doctor/SessionStart. Use for "export to AGENTS.md", "sync my agent rules", "fan out the floor", "/hippo:export-agents".
Migration on-ramp — import existing rules/notes from other tools into ranked, staleness-tracked, deduped hippo memories, per-item confirmed and secret-linted. First adapter is Cursor (.cursor/rules/*.mdc, globs become cited paths). Use for "import my cursor rules", "migrate from cursor", "bring in my existing rules", "/hippo:import".
Run this once per NEW project to seed .claude/memory/ from the starter packs, a skeleton MEMORY.md floor, the cross-machine symlink, and a built recall index. Also safe on an EXISTING corpus (teammate clone, new worktree, second machine) — skips seeding, just wires up this machine's symlink+index. Use when a user says "init memory here", "set up memory for this project", "/hippo:init", or opens a fresh/cloned repo. Idempotent — never overwrites an existing memory file.
Create a new, recall-ready memory file the right way (correct frontmatter, backfilled citation provenance, refreshed index, and a floor pointer if applicable). Use whenever the agent decides to save something to memory — a user preference, a corrected mistake, a project fact, or an external reference — instead of hand-writing a markdown file. Triggers include "remember this", "save this to memory", "/hippo:new".
Share or adopt memory packs. Extract chosen corpus memories into a shareable pack dir (manifest.json in the shipped shape, portability-linted); INSTALL a pack from a local dir or cloned repo per-item on the trust spine (consent shows what injects, secret-lint hard gate); UPDATE installed packs with per-item three-way merges preserving local edits. Use for "make a pack", "install this pack", "update the pack from upstream", "/hippo:pack".
Lift ONE proven-portable memory from this project's corpus into the machine-local user tier (or this repo's private tier) with an origin stamp — "learned in <repo>@<sha>" — so the lesson recalls in every other project WITH provenance. Per-item and agent-gated, never bulk. Use for "promote this memory", "make this lesson global", "lift to the user tier", "/hippo:promote".
Promote ONE reinforced procedural memory into a glob-SCOPED .claude/rules/<name>.md — the harness lazy-loads it only for edits under paths derived from the memory's cited_paths, instead of another always-load CLAUDE.md line. Propose-only reviewable diff, over-scoping-capped so a rule never becomes near-unscoped; never a write without your yes. Use for "promote this to a rule", "make this a scoped rule", "scope this lesson to its files", "/hippo:promote-rule".
Publish ONE local-only memory INTO this repo's committed subset — the per-item preflight (mechanical refusals, the review gate with entropy ON, advisory receipt) prints the exact git add -f + commit line and stops; the human runs it. Use for "publish this memory", "commit this memory to the repo", "add this memory to the public subset", "/hippo:publish".
Deliberately recall what memory knows — answer "what do you remember about X", "what do you know here", or "why was that injected", and list the corpus by type. Wraps hybrid recall + the link graph into a human-readable listing (name, type, staleness, inbound/outbound links). Triggers include "what do you remember about", "what do you know here", "recall", "/hippo:recall". The read verb between /hippo:new (write) and /hippo:audit (maintenance); not plumbing (that's /hippo:doctor).
Uninstall/offboarding for THIS project — removes the cross-machine symlink so native memory stops injecting the floor, offers to delete the derived index/telemetry dirs, and reports (never deletes) the shared venv/model-cache paths. The committed .claude/memory/ corpus is always left alone, inert in git. Use for "remove hippo from this project", "uninstall memory", "/hippo:remove".
Drain the contradiction inbox — walk every unresolved `contradicts` pair in the memory corpus and render a per-item human verdict (keep one side and supersede the other, scope both, merge, or mark not-conflicting). Nothing auto-picks a winner. Triggers include "resolve contradictions", "contradiction inbox", "resolve memory conflicts", "which memories conflict", "/hippo:resolve". An empty inbox is fine — it just says so.
Review a memory diff like a PR — a zero-LLM packet that op-classifies each touched memory (ADD/UPDATE/SUPERSEDE/ARCHIVE/EDGE), runs the shipped lints scoped to the touched files, and (local only) previews how recall shifts by replaying recent episode queries against base-vs-head shadow indexes. Use for "review this memory PR", "review the memory diff", "what changed in memory", "memory review packet", "/hippo:review". The human still merges — this renders review material, never an approval.
The recall receipt (glass-box) — explain WHY hippo surfaced a memory for a query (winning backend, typed edges, steering, salience) or why it did NOT (the best candidate's sub-floor near-miss score and the floor it missed, an untrusted corpus, or a no-shared-token BM25 miss). Triggers include "why did you recall that", "why was that injected", "why didn't you remember", "recall receipt", "/hippo:why". Read-only; runs the same ranking the hook uses.
MCP servers1
- command
- ${CLAUDE_PLUGIN_ROOT}/bin/hippo
- args
- mcp
Plugin manifests1
{
"name": "hippo",
"description": "Local, git-native memory for Claude Code — your repo is the store, recall is $0/prompt (no LLM, no network), staleness is git-drift (the cited code moved), and team memory ships through review.",
"version": "1.36.1",
"author": {
"name": "youknowfred",
"url": "https://youknowfred.com"
},
"homepage": "https://github.com/youknowfred/hippo",
"repository": "https://github.com/youknowfred/hippo",
"license": "MIT",
"keywords": [
"memory",
"recall",
"agent-memory",
"knowledge-base"
],
"mcpServers": {
"hippo": {
"command": "${CLAUDE_PLUGIN_ROOT}/bin/hippo",
"args": [
"mcp"
]
}
}
}For maintainers
If you maintain this plugin, link to this source-backed listing from your README so users can review its manifest and indexed components.
[hippo on Agent Plugins Marketplace](https://pluginsmp.com/plugins/hippo)