@sacheeperera/hippocampus
v0.1.0
Published
Your agents forget between sessions. Hippocampus gives them a rolling briefing — synthesized daily, injected automatically.
Downloads
10
Maintainers
Readme
Hippocampus
Your agents forget between sessions. This fixes that.
Every time an OpenClaw agent starts a new session, it wakes up blank. No memory of yesterday's conversations, open threads, or promises made. You've seen it: the same questions re-asked, context re-explained, commitments quietly dropped.
Hippocampus gives your agents a rolling briefing that's automatically injected into every session. No tool calls, no manual loading — your agent just knows.
openclaw plugins install hippocampusBefore and After
Without Hippocampus:
"What were we working on?" "Can you remind me of the status?" "I don't have context on that conversation..."
With Hippocampus:
Your agent opens every session already briefed — what's top of mind, what threads are open, what commitments are outstanding, and what happened in recent sessions.
How It Works
Three memory layers, working together:
Daily Notes HIPPOCAMPUS.md MEMORY.md
(raw session logs) -> (14-day rolling -> (permanent curated
synthesis, auto- knowledge, graduated
injected every when patterns prove
session) durable)- Daily notes —
memory/YYYY-MM-DD.mdfiles your agents already write - HIPPOCAMPUS.md — A daily cron synthesizes your agent's recent notes into a rolling briefing. Automatically loaded into every session via
prependSystemContext. - MEMORY.md — When a pattern persists across multiple cycles, promote it to permanent memory with the
hippocampus_graduatetool.
The synthesis is the key part. It doesn't dump raw logs — it synthesizes. The output reads like an ops briefing: what's hot, what's stuck, what's owed, what just happened.
Quick Start
1. Install
openclaw plugins install hippocampusZero config required for single-agent setups. The plugin activates immediately:
- HIPPOCAMPUS.md injected into every session
- Setup and graduation tools registered
- The
hippocampus-synccompanion skill available for cron
2. Set up the daily synthesis
Ask your agent:
"Set up hippocampus for me"
The setup wizard scans your workspace, shows you what the synthesis will cover, and generates configuration. Works via Telegram, Slack, or any channel.
Or via CLI:
openclaw hippocampus setupOr add the cron job directly:
openclaw cron add \
--agent main \
--schedule "0 4 * * *" \
--name "hippocampus-sync" \
--prompt "Run the hippocampus-sync skill"3. Done
Your agent now wakes up every session with full rolling context.
What Gets Synthesized
The daily synthesis produces these sections:
| Section | What it captures | |---------|-----------------| | Top of Mind | 3-5 highest-priority items, rewritten every sync | | Open Threads | Active work streams, persists until resolved | | Commitments | Who owes what to whom, with staleness flags | | Recent Sessions | Rolling conversation log from the last 14 days |
You can add custom sections per agent (e.g., cron_health for a platform engineer, pipeline_status for a sales agent).
Multi-Agent Setup
Running multiple agents? Configure each one's synthesis independently:
{
"plugins": {
"entries": {
"hippocampus": {
"config": {
"agents": {
"cyclawps": {
"domainFraming": "Platform-centric. Lead with system health and infrastructure state.",
"sources": [
{
"id": "health",
"path": "artifacts/reports/health-check/",
"label": "Health reports",
"whatToExtract": "System state, uptime, errors",
"windowed": true
}
],
"outputSections": ["top_of_mind", "open_threads", "commitments", "recent_sessions", "cron_health"]
},
"bobo": {
"domainFraming": "Strategy-focused. Lead with pipeline health, client relationships, and decisions pending.",
"sources": [
{ "id": "charter", "path": "CHARTER.md", "label": "Strategic charter", "whatToExtract": "Current priorities and goal progress", "windowed": false }
]
}
}
}
}
}
}
}Agent sources are additive — memory/ is always included. Additional sources are appended alongside it.
Each agent gets a hippocampus-sync.config.md companion file in their workspace — human-readable, directly editable — controlling domain framing, source tables, and output sections.
{
"plugins": {
"entries": {
"hippocampus": {
"config": {
"enabled": true,
"rollingWindowDays": 14,
"targetSizeChars": { "min": 3000, "max": 5000 },
"contextInjection": {
"enabled": true,
"priority": 50,
"excludeAgents": []
},
"synthesis": {
"outputSections": ["top_of_mind", "open_threads", "commitments", "recent_sessions"],
"domainFraming": "General-purpose working memory"
},
"graduation": {
"enabled": true,
"persistenceDaysThreshold": 14,
"autoSuggest": true,
"maxMemorySizeChars": 15000
},
"agents": {}
}
}
}
}
}| Setting | Default | Description |
|---------|---------|-------------|
| rollingWindowDays | 14 | Days of history the synthesis considers |
| targetSizeChars | 3000-5000 | Size budget for HIPPOCAMPUS.md |
| contextInjection.priority | 50 | Higher = earlier in context window |
| contextInjection.excludeAgents | [] | Agent IDs to skip injection for |
| graduation.persistenceDaysThreshold | 14 | Days a pattern must persist before suggesting graduation |
| graduation.maxMemorySizeChars | 15000 | Size guard for MEMORY.md |
Agent Tools
| Tool | What it does |
|------|-------------|
| hippocampus_setup | Scans your workspace, previews what the synthesis will produce, and generates configuration. Works conversationally — just tell your agent "set up hippocampus." |
| hippocampus_graduate | Promotes a durable learning from the rolling synthesis to permanent MEMORY.md. Deduplicates automatically. |
Coexistence
Hippocampus works alongside other OpenClaw plugins without conflict:
- lossless-claw — Hippocampus is not a context engine. It uses
before_prompt_buildhooks, which lossless-claw respects during context assembly. - memory-core — Hippocampus writes standard markdown files. memory-core indexes them automatically.
- Any plugin — No conflicts. Reads and writes markdown in agent workspaces only.
Development
git clone https://github.com/comicsansbestfont/hippocampus.git
cd hippocampus && npm install
npm test # 47 tests
npx tsc --noEmit # type checkSee docs/DEVELOPMENT.md for the contributor guide and docs/TESTING.md for end-to-end testing with a dev OpenClaw instance.
Background
Built by Sachee Perera. Battle-tested on a 13-agent, 4-business-unit OpenClaw instance running in production.
The three-layer memory pattern emerged from running autonomous agents across GTM advisory, e-commerce, content, and platform engineering — where dropped context means dropped revenue.
License
MIT
