Drop-in memory harness for AI agents — 3-tier memory, compaction tree, hybrid search. One command to set up. Works with Claude Code and OpenClaw.
# Add to your Claude Code skills
git clone https://github.com/kevin-hs-sohn/hipocampusDrop-in memory harness for AI agents. Zero infrastructure — just files.
3-tier memory architecture with a 5-level compaction tree, auto-loaded ROOT.md topic index, and optional hybrid search via qmd. One command to set up, works immediately with Claude Code and OpenClaw bots.
/plugin marketplace add kevin-hs-sohn/hipocampus
/plugin install hipocampus@kevin-hs-sohn/hipocampus
Then run npx hipocampus init for full setup (directories, templates, search, hooks).
npx hipocampus init
This creates the full memory structure in your project:
MEMORY.md # Long-term memory (OpenClaw only — Claude Code uses platform auto memory)
USER.md # User profile (OpenClaw only — Claude Code uses platform auto memory)
SCRATCHPAD.md # Active working state
WORKING.md # Current tasks in progress
TASK-QUEUE.md # Task backlog (queued items only)
memory/ # ROOT.md + daily logs + 5-level compaction tree
knowledge/ # Searchable knowledge base
plans/ # Task plans
hipocampus.config.json # Configuration
.claude/skills/ # Agent skills (hipocampus-core, hipocampus-compaction, hipocampus-search)
It also:
--no-search)No comments yet. Be the first to share your thoughts!
.gitignore# Disable vector search (BM25 only, saves ~2GB disk)
npx hipocampus init --no-vector
# Skip qmd entirely (compaction tree + manual file reads only)
npx hipocampus init --no-search
# Override platform detection (auto-detects by default)
npx hipocampus init --platform claude-code
npx hipocampus init --platform openclaw
Install hipocampus on a Claude Code or OpenClaw project, and your agent gains persistent memory across sessions. It remembers what you worked on, what decisions were made, what lessons were learned — and it knows what it knows without loading everything into context.
The effect is similar to injecting your entire conversation history into every API call, but at a fraction of the token cost (~3K tokens instead of 100K+).
Modern models support 200K–1M token context windows. You could theoretically dump all past history into context — 500K tokens for conversation, 500K for past memory. But this creates two problems: